Development Update

[vladmandic]

Update

This has been pending for a while, but finally uploaded some massive changes

Ask for help

To avoid having this repo rely just on me, I'd love to have additional maintainers with full admin rights. If you're interested, ping me!

Additionally, I'd love area-experts for different platform to help with best practices:

• Running on AMD GPUs
• Running on CPU
• Running on Apple M1
• Running on Windows

The goal is to actually integrate all best options for uniform out-of-the-box experience without the need to tweak anything manually.

How WebUI is started is pretty much completely changed:

This should be fully cross-platform, but I would really love to have additional contibutors and/or maintainers to join and help lead the effords on different platforms

• webui.bat and webui.sh:
  Platform specific wrapper scripts that starts launch.py in Python virtual environment
  Note: Server can run without virtual environment, but it is recommended to use it
  This is carry-over from original repo
  If you're unsure which launcher to use, this is the one you want
• launch.py:
  Main startup script
  Can be used directly to start server in manually activated venv or to run it without venv
• setup.py:
  Main installer, used by launch.py
  Can also be used directly to update repository or extensions
  If running manually, make sure to activate venv first (if used)
  Setup details are logged to setup.log
• webui.py:
  Main server script

Note: by default, launcher/setup will re-check dependencies and update extensions and submodules on each startup.
This can be disabled with cmd flag --noupdate

Alternatively, you can run in quick mode if you're sure everything is up-to-date by adding cmd flag --quick

Btw, if you're wondering what takes most time during startup: its running installers for each extension - so having large number of extensions installed has a major performance impact on server startup time

Example startup sequence:

11:05:33-591557 INFO     Python 3.10.6 on Linux
11:05:33-593285 INFO     Running setup
11:05:33-593987 INFO     Installing requirements
11:05:33-651731 INFO     Version: 81b8294e Wed Apr 12 10:40:11 2023 -0400
11:05:33-945552 WARNING  Latest available version: 2023-04-11T00:12:27Z
11:05:33-949224 INFO     Updating Wiki
11:05:35-694480 INFO     Installing packages
11:05:35-695232 INFO     Installing repositories
11:05:35-700296 INFO     Installing submodules
11:05:35-788475 INFO     Updating submodules
11:05:39-769383 INFO     Built-in extensions: ['stable-diffusion-webui-images-browser', 'sd-webui-model-converter', 'seed_travel', 'sd-extension-aesthetic-scorer', 'clip-interrogator-ext', 'stable-diffusion-webui-rembg', 'sd-extension-system-info', 'SwinIR', 'sd-extension-steps-animation', 'prompt-bracket-checker', 'LDSR', 'Lora', 'sd-dynamic-thresholding', 'sd-webui-controlnet']
11:05:49-397596 INFO     Enabled extensions: ['sd-webui-3d-open-pose-editor', 'openOutpaint-webUI-extension', 'stable-diffusion-webui-inspiration', 'a1111-sd-webui-locon', 'openpose-editor']
11:05:52-896265 INFO     Installing requirements
11:05:52-899582 INFO     Setting environment tuning
11:05:52-900085 INFO     Server Arguments: []
11:05:56-195850 INFO     Torch 2.0.0+cu118 CUDA 11.8 cuDNN 8700
11:05:56-197752 INFO     GPU NVIDIA GeForce RTX 3060 VRAM 12288 Arch (8, 6) Cores 28

New logger and new error and exception handlers

More work to be done, but all essentials are in place

Built-in performance profiler

Enable with --profile flag

Simplified dependencies and updated core libraries

Yes, requirements-versions.txt is gone

Settings changes

Quite a few defaults have been changed, but if you're have your config.json, it will remain as-is
To reset settings to defaults, delete config.json and restart server

UI changes

Until I make CSS cleaner, this is a bit ugly so I do not recommend using 3rd party themes for now

Moved almost all command line params to settings configurable from UI

This includes choices of cross-optimization and memory attention as well as configuration of system paths

See:

• Settings -> System Paths
• Settings -> Stable Diffusion

Yes, cmd flags like --xformers are gone

This is whats left:

python launch.py --help

options:
  -h, --help                           show this help message and exit
  --no-half                            Do not switch the model to 16-bit floats
  --no-half-vae                        Do not switch the VAE model to 16-bit floats
  --precision {full,autocast}          Evaluate at this precision
  --medvram                            Enable model optimizations for sacrificing a little speed for low memory usage
  --lowvram                            Enable model optimizations for sacrificing a lot of speed for lowest memory usage
  --lowram                             Load checkpoint weights to VRAM instead of RAM
  --ckpt CKPT                          Path to checkpoint of stable diffusion model to load immediately
  --vae VAE                            Path to checkpoint of stable diffusion VAE model to load immediately
  --data-dir DATA_DIR                  Base path where all user data is stored
  --allow-code                         Allow custom script execution
  --share                              Enable to make the UI accessible through Gradio site
  --enable-insecure                    Enable extensions tab regardless of other options
  --use-cpu USE_CPU [USE_CPU ...]      Force use CPU for specified modules
  --listen                             Launch web server using public IP address
  --port PORT                          Launch web server with given server port
  --hide-ui-dir-config                 Hide directory configuration from UI
  --freeze-settings                    Disable editing settings
  --gradio-auth GRADIO_AUTH            Set Gradio authentication like "username:password,username:password""
  --gradio-auth-path GRADIO_AUTH_PATH  Set Gradio authentication using file
  --autolaunch                         Open the UI URL in the system's default browser upon launch
  --disable-console-progressbars       Do not output progressbars to console
  --disable-safe-unpickle              Disable checking models for malicious code
  --api-auth API_AUTH                  Set API authentication
  --api-log                            Enable logging of all API requests
  --device-id DEVICE_ID                Select the default CUDA device to use
  --cors-origins CORS_ORIGINS          Allowed CORS origin(s) in the form of a comma-separated list
  --cors-regex CORS_REGEX              Allowed CORS origin(s) in the form of a single regular expression
  --tls-keyfile TLS_KEYFILE            Partially enables TLS, requires --tls-certfile to fully function
  --tls-certfile TLS_CERTFILE          Partially enables TLS, requires --tls-keyfile to fully function
  --server-name SERVER_NAME            Sets hostname of server
  --no-hashing                         Disable sha256 hashing of checkpoints
  --no-download-sd-model               Disable download of default model even if no model is found
  --profile                            Run profiler, default: False
  --debug                              Run installer with debug logging, default: False
  --quick                              Skip installing if setup.log is newer than repo timestamp, default: False
  --upgrade                            Upgrade main repository to latest version, default: False
  --noupdate                           Skip update extensions and submodules, default: False
  --skip-extensions                    Skips running individual extension installers, default: False

List of my TBD items

• Reconnect UI to ops in progress on browser restart
• Redo Extensions tab: see https://vladmandic.github.io/sd-extension-manager/pages/extensions.html
• Cleanup & full integrate CSS style
• Replace PngInfo/EXIF metadata handler
• Pick & merge PRs from main repo
• Create new GitHub hooks/actions for CI/CD
• Investigate integration with Torch-DirectML

[vladmandic]

I've just finished major update to installer/launcher as well as some environment auto-tunes.
Hopefully now those steps should not be needed and should work out-of-the-box

Check README for updated notes and updated thread #92 with details.

[vladmandic]

Update for 04/14/2023:

• handle duplicate extensions
• redo exception handler
• fix generate forever
• enable cmdflags compatibility
• change default css font
• fix ti previews on initial start
• enhance tracebacks
• pin transformers version to last known good version
• fix extension loader

[vladmandic]

Update for 04/15/2023:

• update default values
• remove ui-config.json from repo, its not fully under user control
• updated extensions mangager
• updated locon/lycoris plugin
• enable quick launch by default
• add multidiffusion upscaler extensions
• add model keyword extension
• enable strong linting
• fix circular imports
• fix extensions updated
• fix git update issues
• update github templates

[vladmandic]

Update for 04/16/2023:

• support for ui themes! to to settings -> user interface -> "ui theme*
  includes 12 predefined themes
• ability to restart server from ui
• updated requirements
• removed styles.csv from repo, its now fully under user control
• removed model-keyword extension as overly aggresive
• rewrite of the fastapi middleware handlers
• install bugfixes, hopefully new installer is now ok
  i really want to focus on features and not troubleshooting installer

[vladmandic]

Update for 04/17/2023:

• themes are now dynamic and discovered from list of available gradio themes on huggingface
  its quite a list of 30+ supported themes so far
• added option to see theme preview without the need to apply it or restart server
• integrated image info functionality into process image tab and removed separate image info tab
• more installer improvements
• fix urls
• updated github integration
• make model download as optional if no models found

[vladmandic]

Update for 04/18/2023:

• reconnect ui to active session on browser restart
  this is one of most frequently asked for items, finally figured it out
  works for text and image generation, but not for process as there is no progress bar reported there to start with
• force unload xformers when not used
  improves compatibility with AMD/M1 platforms
• add styles.csv to UI settings to allow customizing path
• add --skip-git to cmd flags for power users that want
  to skip all git checks and operations and perform manual updates
• add --disable-queue to cmd flags that disables Gradio queues (experimental)
  this forces it to use HTTP instead of WebSockets and can help on unreliable network connections
• set scripts & extensions loading priority and allow custom priorities
  fixes random extension issues:
  ScuNet upscaler dissapearing, Additional Networks not showing up on XYZ axis, etc.
• improve html loading order
• remove some asserts causing runtime errors and replace with user-friendly messages
• update README.md
• update TODO.md

[vladmandic]

Update for 04/19/2023:

• fix live preview
• fix model merge
• fix handling of user-defined temp folders
• fix submit benchmark
• option to override torch and xformers installer
• separate benchmark data for system-info extension
• minor css fixes
• created initial merge backlog from pending prs on a1111 repo
  see Pending PRs #258 for details

[vladmandic]

Update for 04/20/2023:

• full CUDA tuning section in UI Settings
• improve exif/pnginfo metadata parsing
  it can now handle 3rd party images or images edited in external software
• optimized setup performance and logging
• improve compatibility with some 3rd party extensions
  for example handle extensions that install packages directly from github urls
• fix initial model download if no models found
• fix vae not found issues
• fix multiple git issues

note: if you previously had command line optimizations such as --no-half, those are now ignored and moved to ui settings

[vladmandic]

Update for 04/22/2023:

• autodetect which system libs should be installed
  this is a first pass of autoconfig for nVidia vs AMD environments
• fix parse cmd line args from extensions
• only install xformers if actually selected as desired cross-attention method
• do not attempt to use xformers or sdp if running on cpu
• merge tomesd token merging
• merge 23 PRs pending from a1111 backlog (!!)

expect shorter updates for the next few days as i'll be partically ooo

[vladmandic]

Update for 04/23/2023:

• fix VAE dtype
  should fix most issues with NaN or black images
• add built-in Gradio themes
• reduce requirements
• more AMD specific work
• initial work on Apple platform support
• additional PR merges
• handle torch cuda crashing in setup
• fix setup race conditions
• fix ui lightbox
• mark tensorflow as optional
• add additional image name templates

[vladmandic]

Update for 04/24/2023:

• identify race condition where generate locks up while fetching preview
• add pulldowns to x/y/z script
• add VAE rollback feature in case of NaNs
• use samples format for live preview
• add token merging
• use Approx NN for live preview
• create default styles.csv
• fix setup not installing tensorflow dependencies
• update default git flags to reduce number of warnings

[vladmandic]

Update for 04/25/2023:

• update process image -> info
• add VAE info to metadata
• update GPU utility search paths for better GPU type detection
• update git flags for wider compatibility
• update environment tuning
• update ti training defaults
• update VAE search paths
• add compatibility opts for some old extensions
• validate script args for always-on scripts
  fixes: deforum with controlnet

[vladmandic]

Update for 04/27/2023:

a bit shorter list as:

1. i've been busy with buxfixing
   there are a lot of them, not going to list each here.
   but seems like critical issues backlog is quieting down and soon i can focus on new features development.
2. i've started collaboration with couple of major projects,
   hopefully this will accelerate future development.

• ability to view/add/edit model description shown in extra networks cards
• add option to specify fallback sampler if primary sampler is not compatible with desired operation
• make clip skip a local parameter
• remove obsolete items from UI settings
• set defaults for AMD ROCm
  if you have issues, you may want to start with a fresh install so configuration can be created from scratch
• set defaults for Apple M1/M2
  if you have issues, you may want to start with a fresh install so configuration can be created from scratch

[vladmandic]

Update for 05/01/2023

Been a bit quieter for last few days as changes were quite significant, but finally here we are...

• Updated core libraries: Gradio, Diffusers, Transformers
• Added support for Intel ARC GPUs via Intel OneAPI IPEX (auto-detected)
• Added support for TorchML (set by default when running on non-compatible GPU or on CPU)
• Enhanced support for AMD GPUs with ROCm
• Enhanced support for Apple M1/M2
• Redesigned command params: run webui --help for details
• Redesigned API and script processing
• Experimental support for multiple Torch compile options
• Improved sampler support
• Google Colab: https://colab.research.google.com/drive/126cDNwHfifCyUpCCQF9IHpEdiXRfHrLN
  Maintained by https://github.com/Linaqruf/sd-notebook-collection
• Fixes, fixes, fixes...

To take advantage of new out-of-the-box tunings, its recommended to delete your config.json so new defaults are applied. Its not necessary, but otherwise you may need to play with UI Settings to get the best of Intel ARC, TorchML, ROCm or Apple M1/M2.

[vladmandic]

Update for 05/04/2023

Again, few days later...

• reviewed/ported all commits from A1111 upstream
  some a few are not applicable as i already have alternative implementations
  and very few i choose not to implement (save/restore last-known-good-config is a bad hack)
  otherwise, we're fully up to date (its doesn't show on fork status as code merges were mostly manual due to conflicts)
  but...due to sheer size of the updates, this may introduce some temporary issues
• redesigned server restart function
  now available and working in ui
  actually, since server restart is now a true restart and not ui restart, it can be used much more flexibly
• faster model load
  plus support for slower devices via stream-load function (in ui settings)
• better logging
  this includes new --debug flag for more verbose logging when troubleshooting

[vladmandic]

Update for 2023-08-20

Another release thats been baking in dev branch for a while...

• general:
  • caching of extra network information to enable much faster create/refresh operations
    thanks @midcoastal
• diffusers:
  • add hires support (experimental)
    applies to all model types that support img2img, including sd and sd-xl
    also supports all hires upscaler types as well as standard params like steps and denoising strength
    when used with sd-xl, it can be used with or without refiner loaded
    how to enable - there are no explicit checkboxes other than second pass itself:
    • hires: upscaler is set and target resolution is not at default
    • refiner: if refiner model is loaded
  • images save options: before hires, before refiner
  • redo move model to cpu logic in settings -> diffusers to be more reliable
    note that system defaults have also changed, so you may need to tweak to your liking
  • update dependencies

[vladmandic]

Update for 2023-08-30

Time for a quite a large update that has been leaking bit-by-bit over the past week or so...
Note: due to large changes, it is recommended to reset (delete) your ui-config.json

• diffusers:
  • support for distilled sd models
    just go to models/huggingface and download a model, for example:
    segmind/tiny-sd, segmind/small-sd, segmind/portrait-finetuned
    those are lower quality, but extremely small and fast
    up to 50% faster than sd 1.5 and execute in as little as 2.1gb of vram
• general:
  • redesigned settings
    • new layout with separated sections:
      settings, ui config, licenses, system info, benchmark, models
    • system info tab is now part of settings
      when running outside of sdnext, system info is shown in main ui
    • all system and image paths are now relative by default
    • add settings validation when performing load/save
    • settings tab in ui now shows settings that are changed from default values
    • settings tab switch to compact view
  • update gradio major version
    this may result in some smaller layout changes since its a major version change
    however, browser page load is now much faster
  • optimizations:
    • optimize model hashing
    • add cli param --skip-all that skips all installer checks
      use at personal discretion, but it can be useful for bulk deployments
    • add model precompile option (when model compile is enabled)
    • extra network folder info caching
      results in much faster startup when you have large number of extra networks
    • faster xyz grid switching
      especially when using different checkpoints
  • update second pass options for clarity
  • models:
    • civitai download missing model previews
  • add openvino (experimental) cpu optimized model compile and inference
    enable with --use-openvino
    thanks @Disty0
  • enable batch img2img scale-by workflows
    now you can batch process with rescaling based on each individual original image size
  • fixes:
    • fix extra networks previews
    • css fixes
    • improved extensions compatibility (e.g. sd-cn-animation)
    • allow changing vae on-the-fly for both original and diffusers backend

[vladmandic]

Update for 2023-09-06

One week later, another large update!

• system:
  • full python 3.11 support
    note that changing python version does require reinstall
    and if you're already on python 3.10, really no need to upgrade
• themes:
  • new default theme: black-teal
  • new light theme: light-teal
  • new additional theme: midnight-barbie, thanks @nyxia
• extra networks:
  • support for tags
    show tags on hover, search by tag, list tags, add to prompt, etc.
  • styles are now also listed as part of extra networks
    existing styles.csv is converted upon startup to individual styles inside models/style
    this is stage one of new styles functionality
    old styles interface is still available, but will be removed in future
  • cache file lists for much faster startup
    speedups are 50+% for large number of extra networks
  • ui refresh button now refreshes selected page, not all pages
  • simplified handling of descriptions
    now shows on-mouse-over without the need for user interaction
  • metadata and info buttons only show if there is actual content
• diffusers:
  • add full support for textual inversions (embeddings)
    this applies to both sd15 and sdxl
    thanks @AI-Casanova for porting compel/sdxl code
  • mix&match base and refiner models (experimental):
    most of those are "because why not" and can result in corrupt images, but some are actually useful
    also note that if you're not using actual refiner model, you need to bump refiner steps
    as normal models are not designed to work with low step count
    and if you're having issues, try setting prompt parser to "fixed attention" as majority of problems
    are due to token mismatches when using prompt attention
    • any sd15 + any sd15
    • any sd15 + sdxl-refiner
    • any sdxl-base + sdxl-refiner
    • any sdxl-base + any sd15
    • any sdxl-base + any sdxl-base
  • ability to interrupt (stop/skip) model generate
  • added aesthetics score setting (for sdxl)
    used to automatically guide unet towards higher pleasing images
    highly recommended for simple prompts
  • added force zeros setting
    create zero-tensor for prompt if prompt is empty (positive or negative)
• general:
  • rembg remove backgrounds support for is-net model
  • settings now show markers for all items set to non-default values
  • metadata refactored how/what/when metadata is added to images
    should result in much cleaner and more complete metadata
  • pre-create all system folders on startup
  • handle model load errors gracefully
  • improved vram reporting in ui
  • improved script profiling (when running in debug mode)

[vladmandic]

Update for 2023-09-13

Started as a mostly a service release with quite a few fixes, but then...
Major changes how hires works as well as support for a very interesting new model Wuerstchen

• tons of fixes
• changes to hires
  • enable non-latent upscale modes (standard upscalers)
  • when using latent upscale, hires pass is run automatically
  • when using non-latent upscalers, hires pass is skipped by default
    enabled using force hires option in ui
    hires was not designed to work with standard upscalers, but i understand this is a common workflow
  • when using refiner, upscale/hires runs before refiner pass
  • second pass can now also utilize full/quick vae quality
  • note that when combining non-latent upscale, hires and refiner output quality is maximum,
    but operations are really resource intensive as it includes: base->decode->upscale->encode->hires->refine
  • all combinations of: decode full/quick + upscale none/latent/non-latent + hires on/off + refiner on/off
    should be supported, but given the number of combinations, issues are possible
  • all operations are captured in image medata
• diffusers:
  • allow loading of sd/sdxl models from safetensors without online connectivity
  • support for new model: wuerstchen
    its a high-resolution model (1024px+) thats ~40% faster than sd-xl with a bit lower resource requirements
    go to models -> huggingface -> search "warp-ai/wuerstchen" -> download
    its nearly 12gb in size, so be patient :)
• minor re-layout of the main ui
• updated ui hints
• updated models -> civitai
  • search and download loras
  • find previews for already downloaded models or loras
• new option inference mode
  • default is standard torch.no_grad
    new option is torch.inference_only which is slightly faster and uses less vram, but only works on some gpus
• new cmdline param --no-metadata
  skips reading metadata from models that are not already cached
• updated gradio
• styles support for subfolders
• css optimizations
• clean-up logging
  • capture system info in startup log
  • better diagnostic output
  • capture extension output
  • capture ldm output
  • cleaner server restart
  • custom exception handling

[vladmandic]

Update for 2023-10-17

This is a major release, with many changes and new functionality...

Changelog is massive, but do read through or you'll be missing on some very cool new functionality
or even free speedups and quality improvements (regardless of which workflows you're using)!

Note that for this release its recommended to perform a clean install (e.g. fresh git clone)
Upgrades are still possible and supported, but clean install is recommended for best experience
(primary issue with upgrades is not configuration (you can carry over your config.json if desired),
its handling of git submodule which need reinitialization)

• UI
  • added change log to UI
    see System -> Changelog
  • converted submenus from checkboxes to accordion elements
    any ui state including state of open/closed menus can be saved as default!
    see System -> User interface -> Set menu states
  • new built-in theme invoked
    thanks @BinaryQuantumSoul
  • add compact view option in settings -> user interface
  • small visual indicator bottom right of page showing internal server job state
• Extra networks:
  • Details
    • new details interface to view and save data about extra networks
      main ui now has a single button on each en to trigger details view
    • details view includes model/lora metadata parser!
    • details view includes civitai model metadata!
  • Metadata:
    • you can scan civitai
      for missing metadata and previews directly from extra networks
      simply click on button in top-right corner of extra networks page
  • Styles
    • save/apply icons moved to extra networks
    • can be edited in details view
    • support for single or multiple styles per json
    • support for embedded previews
    • large database of art styles included by default
      can be disabled in settings -> extra networks -> show built-in
    • styles can also be used in a prompt directly: <style:style_name>
      if style if an exact match, it will be used
      otherwise it will rotate between styles that match the start of the name
      that way you can use different styles as wildcards when processing batches
    • styles can have extra fields, not just prompt and negative prompt
      for example: "Extra: sampler: Euler a, width: 480, height: 640, steps: 30, cfg scale: 10, clip skip: 2"
  • VAE
    • VAEs are now also listed as part of extra networks
    • Image preview methods have been redesigned: simple, approximate, taesd, full
      please set desired preview method in settings
    • both original and diffusers backend now support "full quality" setting
      if you desired model or platform does not support FP16 and/or you have a low-end hardware and cannot use FP32
      you can disable "full quality" in advanced params and it will likely reduce decode errors (infamous black images)
  • LoRA
    • LoRAs are now automatically filtered based on compatibility with currently loaded model
      note that if lora type cannot be auto-determined, it will be left in the list
  • Refiner
    • you can load model from extra networks as base model or as refiner
      simply select button in top-right of models page
  • General
    • faster search, ability to show/hide/sort networks
    • refactored subfolder handling
      note: this will trigger model hash recaclulation on first model use
• Diffusers:
  • better pipeline auto-detect when loading from safetensors
  • SDXL Inpaint
    • although any model can be used for inpainiting, there is a case to be made for
      dedicated inpainting models as they are tuned to inpaint and not generate
    • model can be used as base model for img2img or refiner model for txt2img
      To download go to Models -> Huggingface:
      • diffusers/stable-diffusion-xl-1.0-inpainting-0.1 (6.7GB)
  • SDXL Instruct-Pix2Pix
    • model can be used as base model for img2img or refiner model for txt2img
      this model is massive and requires a lot of resources!
      to download go to Models -> Huggingface:
      • diffusers/sdxl-instructpix2pix-768 (11.9GB)
  • SD Latent Upscale
    • you can use SD Latent Upscale models as refiner models
      this is a bit experimental, but it works quite well!
      to download go to Models -> Huggingface:
      • stabilityai/sd-x2-latent-upscaler (2.2GB)
      • stabilityai/stable-diffusion-x4-upscaler (1.7GB)
  • better Prompt attention
    should better handle more complex prompts
    for sdxl, choose which part of prompt goes to second text encoder - just add TE2: separator in the prompt
    for hires and refiner, second pass prompt is used if present, otherwise primary prompt is used
    new option in settings -> diffusers -> sdxl pooled embeds
    thanks @AI-Casanova
  • better Hires support for SD and SDXL
  • better TI embeddings support for SD and SDXL
    faster loading, wider compatibility and support for embeddings with multiple vectors
    information about used embedding is now also added to image metadata
    thanks @AI-Casanova
  • better Lora handling
    thanks @AI-Casanova
  • better SDXL preview quality (approx method)
    thanks @BlueAmulet
  • new setting: settings -> diffusers -> force inpaint
    as some models behave better when in inpaint mode even for normal img2img tasks
• Upscalers:
  • pretty much a rewrite and tons of new upscalers - built-in list is now at 42
  • fix long outstanding memory leak in legacy code, amazing this went undetected for so long
  • more high quality upscalers available by default
    SwinIR (2), ESRGAN (12), RealESRGAN (6), SCUNet (2)
  • if that is not enough, there is new chaiNNer integration:
    adds 15 more upscalers from different families out-of-the-box:
    HAT (6), RealHAT (2), DAT (1), RRDBNet (1), SPSRNet (1), SRFormer (2), SwiftSR (2)
    and yes, you can download and add your own, just place them in models/chaiNNer
  • two additional latent upscalers based on SD upscale models when using Diffusers backend
    SD Upscale 2x, SD Upscale 4x*
    note: Recommended usage for SD Upscale is by using second pass instead of upscaler
    as it allows for tuning of prompt, seed, sampler settings which are used to guide upscaler
  • upscalers are available in xyz grid
  • simplified settings->postprocessing->upscalers
    e.g. all upsamplers share same settings for tiling
  • allow upscale-only as part of txt2img and img2img workflows
    simply set denoising strength to 0 so hires does not get triggered
  • unified init/download/execute/progress code
  • easier installation
• Samplers:
  • moved ui options to submenu
  • default list for new installs is now all samplers, list can be modified in settings
  • simplified samplers configuration in settings
    plus added few new ones like sigma min/max which can highly impact sampler behavior
  • note that list of samplers is now different since keeping a flat-list of all possible
    combinations results in 50+ samplers which is not practical
    items such as algorithm (e.g. karras) is actually a sampler option, not a sampler itself
• CivitAI:
  • civitai model download is now multithreaded and resumable
    meaning that you can download multiple models in parallel
    as well as resume aborted/incomplete downloads
  • civitai integration in models -> civitai can now find most
    previews AND metadata for most models (checkpoints, loras, embeddings)
    metadata is now parsed and saved in [model].json
    typical hit rate is >95% for models, loras and embeddings
  • description from parsed model metadata is used as model description if there is no manual
    description file present in format of [model].txt
  • to enable search for models, make sure all models have set hash values
    Models -> Valida -> Calculate hashes
• LoRA
  • new unified LoRA handler for all LoRA types (lora, lyco, loha, lokr, locon, ia3, etc.)
    applies to both original and diffusers backend
    thanks @AI-Casanova for diffusers port
  • for backend:original, separate lyco handler has been removed
• Compute
  • CUDA:
    • default updated to torch 2.1.0 with cuda 12.1
    • testing moved to torch 2.2.0-dev/cu122
    • check out generate context menu -> show nvml for live gpu stats (memory, power, temp, clock, etc.)
  • Intel Arc/IPEX:
    • tons of optimizations, built-in binary wheels for Windows
      i have to say, intel arc/ipex is getting to be quite a player, especially with openvino
      thanks @Disty0 @Nuullll
  • AMD ROCm:
    • updated installer to support detect ROCm 5.4/5.5/5.6/5.7
    • support for torch-rocm-5.7
  • xFormers:
    • default updated to 0.0.23
    • note that latest xformers are still not compatible with cuda 12.1
      recommended to use torch 2.1.0 with cuda 11.8
      if you attempt to use xformers with cuda 12.1, it will force a full xformers rebuild on install
      which can take a very long time and may/may-not work
    • added cmd param --use-xformers to force usage of exformers
  • GC:
    • custom garbage collect threshold to reduce vram memory usage, thanks @Disty0
      see settings -> compute -> gc
• Inference
  • new section in settings
    • HyperTile: new!
      available for diffusers and original backends
      massive (up to 2x) speed-up your generations for free :)
      note: hypertile is not compatible with any extension that modifies processing parameters such as resolution
      thanks @tfernd
    • Free-U: new!
      available for diffusers and original backends
      improve generations quality at no cost (other than finding params that work for you)
      note: temporarily disabled for diffusers pending release of diffusers==0.22
      thanks @ljleb
    • Token Merging: not new, but updated
      available for diffusers and original backends
      speed-up your generations by merging redundant tokens
      speed up will depend on how aggressive you want to be with token merging
    • Batch mode
      new option settings -> inference -> batch mode
      when using img2img process batch, optionally process multiple images in batch in parallel
      thanks @Symbiomatrix
• NSFW Detection/Censor
  • install extension: NudeNet
    body part detection, image metadata, advanced censoring, etc...
    works for text, image and process workflows
    more in the extension notes
• Extensions
  • automatic discovery of new extensions on github
    no more waiting for them to appear in index!
  • new framework for extension validation
    extensions ui now shows actual status of extensions for reviewed extensions
    if you want to contribute/flag/update extension status, reach out on github or discord
  • better overall compatibility with A1111 extensions (up to a point)
  • MultiDiffusion
    has been removed from list of built-in extensions
    you can still install it manually if desired
  • [LyCORIS]https://github.com/KohakuBlueleaf/a1111-sd-webui-lycoris
    has been removed from list of built-in extensions
    it is considered obsolete given that all functionality is now built-in
• General
  • Startup
    • all main CLI parameters can now be set as environment variable as well
      for example --data-dir <path> can be specified as SD_DATADIR=<path> before starting SD.Next
  • XYZ Grid
    • more flexibility to use selection or strings
  • Logging
    • get browser session info in server log
    • allow custom log file destination
      see webui --log
    • when running with --debug flag, log is force-rotated
      so each sdnext.log.* represents exactly one server run
    • internal server job state tracking
  • Launcher
    • new webui.ps1 powershell launcher for windows (old webui.bat is still valid)
      thanks @em411
  • API
    • add end-to-end example how to use API: cli/simple-txt2img.js
      covers txt2img, upscale, hires, refiner
  • train.py
    • wrapper script around built-in kohya's lora training script
      see cli/train.py --help
      new support for sd and sdxl, thanks @evshiron
      new support for full offline mode (without sdnext server running)

[vladmandic]

Update for 2023-11-06

Another pretty big release, this time with focus on new models (3 new model types), new backends and optimizations
Plus quite a few fixes

Also, Wiki has been updated with new content, so check it out!
Some highlights: OpenVINO, IntelArc, DirectML, ONNX/Olive

• Diffusers
  • since now SD.Next supports 12 different model types, we've added reference model for each type in
    Extra networks -> Reference for easier select & auto-download
    Models can still be downloaded manually, this is just a convenience feature & a showcase for supported models
  • new model type: Segmind SSD-1B
    its a distilled model trained at 1024px, this time 50% smaller and faster version of SD-XL!
    (and quality does not suffer, its just more optimized)
    test shows batch-size:4 with 1k images at full quality used less than 6.5GB of VRAM
    and for further optimization, you can use built-in TAESD decoder,
    which results in batch-size:16 with 1k images using 7.9GB of VRAM
    select from extra networks -> reference or download using built-in Huggingface downloader: segmind/SSD-1B
  • new model type: Pixart-α XL 2
    in medium/512px and large/1024px variations
    comparable in quality to SD 1.5 and SD-XL, but with better text encoder and highly optimized training pipeline
    so finetunes can be done in as little as 10% compared to SD/SD-XL (note that due to much larger text encodeder, it is a large model)
    select from extra networks -> reference or download using built-in Huggingface downloader: PixArt-alpha/PixArt-XL-2-1024-MS
  • new model type: LCM: Latent Consistency Models
    trained at 512px, but with near-instant generate in a as little as 3 steps!
    combined with OpenVINO, generate on CPU takes less than 5-10 seconds: https://www.youtube.com/watch?v=b90ESUTLsRo
    and absolute beast when combined with HyperTile and TAESD decoder resulting in 28 FPS
    (on RTX4090 for batch 16x16 at 512px)
    note: set sampler to Default before loading model as LCM comes with its own LCMScheduler sampler
    select from extra networks -> reference or download using built-in Huggingface downloader: SimianLuo/LCM_Dreamshaper_v7
  • support for Custom pipelines, thanks @Disty0
    download using built-in Huggingface downloader
    think of them as plugins for diffusers not unlike original extensions that modify behavior of ldm backend
    list of community pipelines: https://github.com/huggingface/diffusers/blob/main/examples/community/README.md
  • new custom pipeline: Disty0/zero123plus-pipeline, thanks @Disty0
    generate 4 output images with different camera positions: front, side, top, back!
    for more details, see Zero123++ #2421
  • new backend: ONNX/Olive (experimental), thanks @lshqqytiger
    for details, see WiKi
  • extend support for Free-U
    improve generations quality at no cost (other than finding params that work for you)
• General
  • attempt to auto-fix invalid samples which occure due to math errors in lower precision
    example: RuntimeWarning: invalid value encountered in cast: sample = sample.astype(np.uint8)
    begone black images (note: if it proves as working, this solution will need to be expanded to cover all scenarios)
  • add Lora OFT support, thanks @antis0007 and @AI-Casanova
  • Upscalers
    • compile option, thanks @Disty0
    • chaiNNer add high quality models from Helaman
  • redesigned Progress bar with full details on current operation
  • new option: settings -> images -> keep incomplete
    can be used to skip vae decode on aborted/skipped/interrupted image generations
  • new option: settings -> system paths -> models
    can be used to set custom base path for all models (previously only as cli option)
  • remove external clone of items in /repositories
  • Interrogator module has been removed from extensions-builtin
    and fully implemented (and improved) natively
• UI
  • UI tweaks for default themes
  • UI switch core font in default theme to noto-sans
    previously default font was simply system-ui, but it lead to too much variations between browsers and platforms
  • UI tweaks for mobile devices, thanks @iDeNoh
  • updated Context menu
    right-click on any button in action menu (e.g. generate button)
• Extra networks
  • sort by name, size, date, etc.
  • switch between gallery and list views
  • add tags from user metadata (in addition to tags in model metadata) for lora
  • added Reference models for diffusers backend
  • faster enumeration of all networks on server startup
• Packages
  • updated diffusers to 0.22.0, transformers to 4.34.1
  • update openvino, thanks @Disty0
  • update directml, @lshqqytiger
• Compute
  • OpenVINO:
    • updated to mainstream torch 2.1.0
    • support for ESRGAN upscalers
• Fixes
  • fix freeu for backend original and add it to xyz grid
  • fix loading diffuser models in huggingface format from non-standard location
  • fix default styles looking in wrong location
  • fix missing upscaler folder on initial startup
  • fix handling of relative path for models
  • fix simple live preview device mismatch
  • fix batch img2img
  • fix diffusers samplers: dpm++ 2m, dpm++ 1s, deis
  • fix new style filename template
  • fix image name template using model name
  • fix image name sequence
  • fix model path using relative path
  • fix safari/webkit layour, thanks @eadnams22
  • fix torch-rocm and tensorflow-rocm version detection, thanks @xangelix
  • fix chainner upscalers color clipping
  • fix for base+refiner workflow in diffusers mode: number of steps, diffuser pipe mode
  • fix for prompt encoder with refiner in diffusers mode
  • fix prompts-from-file saving incorrect metadata
  • fix add/remove extra networks to prompt
  • fix before-hires step
  • fix diffusers switch from invalid model
  • force second requirements check on startup
  • remove lyco, multiple_tqdm
  • enhance extension compatibility for exensions directly importing codeformers
  • enhance extension compatibility for exensions directly accessing processing params
  • css fixes
  • clearly mark external themes in ui
  • update typing-extensions

[vladmandic]

Update for 2023-11-23

New release, primarily focused around three major new features: full LCM support, completely new Model Merge functionality and Stable-fast compile support
Also included are several other improvements and large number of hotfixes - see full changelog for details

• Diffusers
  • LCM support for any SD 1.5 or SD-XL model!
    • download lcm-lora-sd15 and/or lcm-lora-sdxl
    • load for favorite SD 1.5 or SD-XL model (original LCM was SD 1.5 only, this is both)
    • load lcm lora (note: lcm lora is processed differently than any other lora)
    • set sampler to LCM
    • set number of steps to some low number, for SD-XL 6-7 steps is normally sufficient
      note: LCM scheduler does not support steps higher than 50
    • set CFG to between 1 and 2
  • Add cli/lcm-convert.py script to convert any SD 1.5 or SD-XL model to LCM model
    by baking in LORA and uploading to Huggingface, thanks @Disty0
  • Support for Stable Fast model compile on Windows/Linux/WSL2 with CUDA
    See Wiki:Benchmark for details and comparisment
    of different backends, precision modes, advanced settings and compile modes
    Hint: 70+ it/s is possible on RTX4090 with no special tweaks
  • Add additional pipeline types for manual model loads when loading from safetensors
  • Updated logic for calculating steps when using base/hires/refiner workflows
  • Improve model offloading for both model and sequential cpu offload when dealing with meta tensors
  • Safe model offloading for non-standard models
  • Fix DPM SDE scheduler
  • Better support for SD 1.5 inpainting models
  • Add support for OpenAI Consistency decoder VAE
  • Enhance prompt parsing with long prompts and support for BREAK keyword
    Change-in-behavior: new line in prompt now means BREAK
  • Add alternative Lora loading algorithm, triggered if SD_LORA_DIFFUSERS is set
• Models
  • Model merge
    • completely redesigned, now based on best-of-class meh by @s1dlx
      and heavily modified for additional functionality and fully integrated by @AI-Casanova (thanks!)
    • merge SD or SD-XL models using simple merge (12 methods),
      using one of presets (20 built-in presets) or custom block merge values
    • merge with ReBasin permuatations and/or clipping protection
    • fully multithreaded for fastest merge possible
  • Model update
    • under UI -> Models - Update
    • scan existing models for updated metadata on CivitAI and
      provide download functionality for models with available
• Extra networks
  • Use multi-threading for 5x load speedup
  • Better Lora trigger words support
  • Auto refresh styles on change
• General
  • Many mobile UI optimizations, thanks @iDeNoh
  • Support for Torch 2.1.1 with CUDA 12.1 or CUDA 11.8
  • Configurable location for HF cache folder
    Default is standard ~/.cache/huggingface/hub
  • Reworked parser when pasting previously generated images/prompts
    includes all txt2img, img2img and override params
  • Reworked model compile
  • Support custom upscalers in subfolders
  • Add additional image info when loading image in process tab
  • Better file locking when sharing config and/or models between multiple instances
  • Handle custom API endpoints when using auth
  • Show logged in user in log when accessing via UI and/or API
  • Support --ckpt none to skip loading a model
• XYZ grid
  • Add refiner options to XYZ Grid
  • Add option to create only subgrids in XYZ grid, thanks @midcoastal
  • Allow custom font, background and text color in settings
• Fixes
  • Fix params.txt saved before actual image
  • Fix inpaint
  • Fix manual grid image save
  • Fix img2img init image save
  • Fix upscale in txt2img for batch counts when no hires is used
  • More uniform models paths
  • Safe scripts callback execution
  • Improved extension compatibility
  • Improved BF16 support
  • Match previews for reference models with downloaded models

[vladmandic]

Update for 2023-12-04

What's new? Native video in SD.Next via both AnimateDiff and Stable-Video-Diffusion - and including native MP4 encoding and smooth video outputs out-of-the-box, not just animated-GIFs.
Also new is support for SDXL-Turbo as well as new Kandinsky 3 models and cool latent correction via HDR controls for any txt2img workflows, best-of-class SDXL model merge using full ReBasin methods and further mobile UI optimizations.

• Diffusers
  • IP adapter
    • lightweight native implementation of T2I adapters which can guide generation towards specific image style
    • supports most T2I models, not limited to SD 1.5
    • models are auto-downloaded on first use
    • for IP adapter support in Original backend, use standard ControlNet extension
  • AnimateDiff
    • lightweight native implementation of AnimateDiff models:
      AnimateDiff 1.4, 1.5 v1, 1.5 v2, AnimateFace
    • supports SD 1.5 only
    • models are auto-downloaded on first use
    • for video saving support, see video support section
    • can be combined with IP-Adapter for even better results!
    • for AnimateDiff support in Original backend, use standard AnimateDiff extension
  • HDR latent control, based on article
    • in Advanced params
    • allows control of latent clamping, color centering and range maximimization
    • supported by XYZ grid
  • SD21 Turbo and SDXL Turbo support
    • just set CFG scale (0.0-1.0) and steps (1-3) to a very low value
    • compatible with original StabilityAI SDXL-Turbo or any of the newer merges
    • download safetensors or select from networks -> reference
  • Stable Video Diffusion and Stable Video Diffusion XT support
    • download using built-in model downloader or simply select from networks -> reference
      support for manually downloaded safetensors models will be added later
    • for video saving support, see video support section
    • go to image tab, enter input image and select script -> stable video diffusion
  • Kandinsky 3 support
    • download using built-in model downloader or simply select from networks -> reference
    • this model is absolutely massive at 27.5GB at fp16, so be patient
    • model params count is at 11.9B (compared to SD-XL at 3.3B) and its trained on mixed resolutions from 256px to 1024px
    • use either model offload or sequential cpu offload to be able to use it
  • better autodetection of inpaint and instruct pipelines
  • support long seconary prompt for refiner
• Video support
  • applies to any model that supports video generation, e.g. AnimateDiff and StableVideoDiffusion
  • support for animated-GIF, animated-PNG and MP4
  • GIF and PNG can be looped
  • MP4 can have additional padding at the start/end as well as motion-aware interpolated frames for smooth playback
    interpolation is done using RIFE with native implementation in SD.Next
    And its fast - interpolation from 16 frames with 10x frames to target 160 frames results takes 2-3sec
  • output folder for videos is in settings -> image paths -> video
• General
  • redesigned built-in profiler
    • now includes both python and torch and traces individual functions
    • use with --debug --profile
  • model merge add SD-XL ReBasin support, thanks @AI-Casanova
  • further UI optimizations for mobile devices, thanks @iDeNoh
  • log level defaults to info for console and debug for log file
  • better prompt display in process tab
  • increase maximum lora cache values
  • fix extra networks sorting
  • fix controlnet compatibility issues in original backend
  • fix img2img/inpaint paste params
  • fix save text file for manually saved images
  • fix python 3.9 compatibility issues

[vladmandic]

Update for 2023-12-29 - Happy holidays everyone!

To wrap up this amazing year, we're releasing a new version of SD.Next, this one is absolutely massive!

Highlights

• Brand new Control module for text, image, batch and video processing
  Native implementation of all control methods for both SD15 and SD-XL
  ▹ ControlNet | ControlNet XS | Control LLLite | T2I Adapters | IP Adapters
  For details, see Wiki documentation:
• Support for new models types out-of-the-box
  This brings number of supported t2i/i2i model families to 13!
  ▹ Stable Diffusion 1.5/2.1 | SD-XL | LCM | Segmind | Kandinsky | Pixart-α | Würstchen | aMUSEd | DeepFloyd IF | UniDiffusion | SD-Distilled | BLiP Diffusion | etc.
• New video capabilities:
  ▹ AnimateDiff | SVD | ModelScope | ZeroScope
• Enhanced platform support
  ▹ Windows | Linux | MacOS with nVidia | AMD | IntelArc | DirectML | OpenVINO | ONNX+Olive backends
• Better onboarding experience (first install)
  with all model types available for single click download & load (networks -> reference)
• Performance optimizations!
  For comparisment of different processing options and compile backends, see Wiki
  As a highlight, we're reaching ~100 it/s (no tricks, this is with full features enabled and end-to-end on a standard nVidia RTX4090)
• New custom pipelines framework for quickly porting any new pipeline

And others improvements in areas such as: Upscaling (up to 8x now with 40+ available upscalers), Inpainting (better quality), Prompt scheduling, new Sampler options, new LoRA types, additional UI themes, better HDR processing, built-in Video interpolation, parallel Batch processing, etc.

Plus some nifty new modules such as FaceID automatic face guidance using embeds during generation and Depth 3D image to 3D scene

Links: Home | Readme | Changelog | Wiki | Discord

Full changelog

• Control
  • native implementation of all image control methods:
    ControlNet, ControlNet XS, Control LLLite, T2I Adapters and IP Adapters
  • top-level Control next to Text and Image generate
  • supports all variations of SD15 and SD-XL models
  • supports Text, Image, Batch and Video processing
  • for details and list of supported models and workflows, see Wiki documentation:
    https://github.com/vladmandic/automatic/wiki/Control
• Diffusers
  • Segmind Vega model support
    • small and fast version of SDXL, only 3.1GB in size!
    • select from networks -> reference
  • aMUSEd 256 and aMUSEd 512 model support
    • lightweigt models that excel at fast image generation
    • note: must select: settings -> diffusers -> generator device: unset
    • select from networks -> reference
  • Playground v1, Playground v2 256, Playground v2 512, Playground v2 1024 model support
    • comparable to SD15 and SD-XL, trained from scratch for highly aesthetic images
    • simply select from networks -> reference and use as usual
  • BLIP-Diffusion
    • img2img model that can replace subjects in images using prompt keywords
    • download and load by selecting from networks -> reference -> blip diffusion
    • in image tab, select blip diffusion script
  • DemoFusion run your SDXL generations at any resolution!
    • in Text tab select script -> demofusion
    • note: GPU VRAM limits do not automatically go away so be careful when using it with large resolutions
      in the future, expect more optimizations, especially related to offloading/slicing/tiling,
      but at the moment this is pretty much experimental-only
  • AnimateDiff
    • overall improved quality
    • can now be used with second pass - enhance, upscale and hires your videos!
  • IP Adapter
    • add support for ip-adapter-plus_sd15, ip-adapter-plus-face_sd15 and ip-adapter-full-face_sd15
    • can now be used in xyz-grid
  • Text-to-Video
    • in text tab, select text-to-video script
    • supported models: ModelScope v1.7b, ZeroScope v1, ZeroScope v1.1, ZeroScope v2, ZeroScope v2 Dark, Potat v1
      if you know of any other t2v models youd like to see supported, let me know!
    • models are auto-downloaded on first use
    • note: current base model will be unloaded to free up resources
  • Prompt scheduling now implemented for Diffusers backend, thanks @AI-Casanova
  • Custom pipelines contribute by adding your own custom pipelines!
    • for details, see fully documented example:
      https://github.com/vladmandic/automatic/blob/dev/scripts/example.py
  • Schedulers
    • add timesteps range, changing it will make scheduler to be over-complete or under-complete
    • add rescale betas with zero SNR option (applicable to Euler, Euler a and DDIM, allows for higher dynamic range)
  • Inpaint
    • improved quality when using mask blur and padding
  • UI
    • 3 new native UI themes: orchid-dreams, emerald-paradise and timeless-beige, thanks @illu_Zn
    • more dynamic controls depending on the backend (original or diffusers)
      controls that are not applicable in current mode are now hidden
    • allow setting of resize method directly in image tab
      (previously via settings -> upscaler_for_img2img)
• Optional
  • FaceID face guidance during generation
    • also based on IP adapters, but with additional face detection and external embeddings calculation
    • calculates face embeds based on input image and uses it to guide generation
    • simply select from scripts -> faceid
    • experimental module: requirements must be installed manually:

      pip install insightface ip_adapter

  • Depth 3D image to 3D scene
    • delivered as an extension, install from extensions tab
      https://github.com/vladmandic/sd-extension-depth3d
    • creates fully compatible 3D scene from any image by using depth estimation
      and creating a fully populated mesh
    • scene can be freely viewed in 3D in the UI itself or downloaded for use in other applications
  • ONNX/Olive
    • major work continues in olive branch, see wiki for details, thanks @lshqqytiger
      as a highlight, 4-5 it/s using DirectML on AMD GPU translates to 23-25 it/s using ONNX/Olive!
• General
  • new onboarding
    • if no models are found during startup, app will no longer ask to download default checkpoint
      instead, it will show message in UI with options to change model path or download any of the reference checkpoints
    • extra networks -> models -> reference section is now enabled for both original and diffusers backend
  • support for Torch 2.1.2 (release) and Torch 2.3 (dev)
  • Process create videos from batch or folder processing
    supports GIF, PNG and MP4 with full interpolation, scene change detection, etc.
  • LoRA
    • add support for block weights, thanks @AI-Casanova
      example <lora:SDXL_LCM_LoRA:1.0:in=0:mid=1:out=0>
    • add support for LyCORIS GLora networks
    • add support for LoRA PEFT (Diffusers) networks
    • add support for Lora-OFT (Kohya) and Lyco-OFT (Kohaku) networks
    • reintroduce alternative loading method in settings: lora_force_diffusers
    • add support for lora_fuse_diffusers if using alternative method
      use if you have multiple complex loras that may be causing performance degradation
      as it fuses lora with model during load instead of interpreting lora on-the-fly
  • CivitAI downloader allow usage of access tokens for download of gated or private models
  • Extra networks new settting -> extra networks -> build info on first access
    indexes all networks on first access instead of server startup
  • IPEX, thanks @Disty0
    • update to Torch 2.1
      if you get file not found errors, set DISABLE_IPEXRUN=1 and run the webui with --reinstall
    • built-in MKL and DPCPP for IPEX, no need to install OneAPI anymore
    • StableVideoDiffusion is now supported with IPEX
    • 8 bit support with NNCF on Diffusers backend
    • fix IPEX Optimize not applying with Diffusers backend
    • disable 32bit workarounds if the GPU supports 64bit
    • add DISABLE_IPEXRUN and DISABLE_IPEX_1024_WA environment variables
    • performance and compatibility improvements
  • OpenVINO, thanks @Disty0
    • 8 bit support for CPUs
    • reduce System RAM usage
    • update to Torch 2.1.2
    • add Directory for OpenVINO cache option to System Paths
    • remove Intel ARC specific 1024x1024 workaround
  • HDR controls
    • batch-aware for enhancement of multiple images or video frames
    • available in image tab
  • Logging
    • additional TRACE logging enabled via specific env variables
      see https://github.com/vladmandic/automatic/wiki/Debug for details
    • improved profiling
      use with --debug --profile
    • log output file sizes
  • Other
    • API several minor but breaking changes to API behavior to better align response fields, thanks @Trojaner
    • Inpaint add option apply_overlay to control if inpaint result should be applied as overlay or as-is
      can remove artifacts and hard edges of inpaint area but also remove some details from original
    • chaiNNer fix NaN issues due to autocast
    • Upscale increase limit from 4x to 8x given the quality of some upscalers
    • Extra Networks fix sort
    • reduced default CFG scale from 6 to 4 to be more out-of-the-box compatibile with LCM/Turbo models
    • disable google fonts check on server startup
    • fix torchvision/basicsr compatibility
    • fix styles quick save
    • add hdr settings to metadata
    • improve handling of long filenames and filenames during batch processing
    • do not set preview samples when using via api
    • avoid unnecessary resizes in img2img and inpaint
    • safe handling of config updates avoid file corruption on I/O errors
    • updated cli/simple-txt2img.py and cli/simple-img2img.py scripts
    • save params.txt regardless of image save status
    • update built-in log monitor in ui, thanks @midcoastal
    • major CHANGELOG doc cleanup, thanks @JetVarimax
    • major INSTALL doc cleanup, thanks JetVarimax

[vladmandic]

Update for 2024-02-07

Another big release just hit the shelves!

Highlights

• A lot more functionality in the Control module:
  • Inpaint and outpaint support, flexible resizing options, optional hires
  • Built-in support for many new processors and models, all auto-downloaded on first use
  • Full support for scripts and extensions
• Complete Face module
  implements all variations of FaceID, FaceSwap and latest PhotoMaker and InstantID
• Much enhanced IPAdapter modules
• Brand new Intelligent masking, manual or automatic
  Using ML models (LAMA object removal, REMBG background removal, SAM segmentation, etc.) and with live previews
  With granular blur, erode and dilate controls
• New models and pipelines:
  Segmind SegMoE, Mixture Tiling, InstaFlow, SAG, BlipDiffusion
• Massive work integrating latest advances with OpenVINO, IPEX and ONNX Olive
• Full control over brightness, sharpness and color shifts and color grading during generate process directly in latent space
• Documentation! This was a big one, with a lot of new content and updates in the WiKi

Plus welcome additions to UI performance, usability and accessibility and flexibility of deployment as well as API improvements
And it also includes fixes for all reported issues so far

As of this release, default backend is set to diffusers as its more feature rich than original and supports many additional models (original backend does remain as fully supported)

Also, previous versions of SD.Next were tuned for balance between performance and resource usage.
With this release, focus is more on performance.
See Benchmark notes for details, but as a highlight, we are now hitting ~110-150 it/s on a standard nVidia RTX4090 in optimal scenarios!

Further details:

• For basic instructions, see README
• For more details on all new features see full CHANGELOG
• For documentation, see WiKi

Full changelog

• Heavily updated Wiki
• Control:
  • new docs:
    • Control overview
    • Control guide, thanks @Aptronymist
  • add inpaint support
    applies to both img2img and controlnet workflows
  • add outpaint support
    applies to both img2img and controlnet workflows
    note: increase denoising strength since outpainted area is blank by default
  • new mask module
    • granular blur (gaussian), erode (reduce or remove noise) and dilate (pad or expand)
    • optional live preview
    • optional auto-segmentation using ml models
      auto-segmentation can be done using segment-anything models or rembg models
      note: auto segmentation will automatically expand user-masked area to segments that include current user mask
    • optional auto-mask
      if you dont provide mask or mask is empty, you can instead use auto-mask to automatically generate mask
      this is especially useful if you want to use advanced masking on batch or video inputs and dont want to manually mask each image
      note: such auto-created mask is also subject to all other selected settings such as auto-segmentation, blur, erode and dilate
    • optional object removal using LaMA model
      remove selected objects from images with a single click
      works best when combined with auto-segmentation to remove smaller objects
    • masking can be combined with control processors in which case mask is applied before processor
    • unmasked part of can is optionally applied to final image as overlay, see settings mask_apply_overlay
  • support for many additional controlnet models
    now built-in models include 30+ SD15 models and 15+ SDXL models
  • allow resize both before and after generate operation
    this allows for workflows such as: image -> upscale or downscale -> generate -> upscale or downscale -> output
    providing more flexibility and than standard hires workflow
    note: resizing before generate can be done using standard upscalers or latent
  • implicit hires
    since hires is only used for txt2img, control reuses existing resize functionality
    any image size is used as txt2img target size
    but if resize scale is also set its used to additionally upscale image after initial txt2img and for hires pass
  • add support for scripts and extensions
    you can now combine control workflow with your favorite script or extension
    note extensions that are hard-coded for txt2img or img2img tabs may not work until they are updated
  • add depth-anything depth map processor and trained controlnet
  • add marigold depth map processor
    this is state-of-the-art depth estimation model, but its quite heavy on resources
  • add openpose xl controlnet
  • add blip/booru interrogate functionality to both input and output images
  • configurable output folder in settings
  • auto-refresh available models on tab activate
  • add image preview for override images set per-unit
  • more compact unit layout
  • reduce usage of temp files
  • add context menu to action buttons
  • move ip-adapter implementation to control tabs
  • resize by now applies to input image or frame individually
    allows for processing where input images are of different sizes
  • support controlnets with non-default yaml config files
  • implement resize modes for override images
  • allow any selection of units
  • dynamically install depenencies required by specific processors
  • fix input image size
  • fix video color mode
  • fix correct image mode
  • fix batch/folder/video modes
  • fix processor switching within same unit
  • fix pipeline switching between different modes
• Face module
  implements all variations of FaceID, FaceSwap and latest PhotoMaker and InstantID
  simply select from scripts and choose your favorite method and model
  note: all models are auto-downloaded on first use
  • FaceID
    • faceid guides image generation given the input image
    • full implementation for SD15 and SD-XL, to use simply select from Scripts
      Base (93MB) uses InsightFace to generate face embeds and OpenCLIP-ViT-H-14 (2.5GB) as image encoder
      Plus (150MB) uses InsightFace to generate face embeds and CLIP-ViT-H-14-laion2B (3.8GB) as image encoder
      SXDL (1022MB) uses InsightFace to generate face embeds and OpenCLIP-ViT-bigG-14 (3.7GB) as image encoder
  • FaceSwap
    • face swap performs face swapping at the end of generation
    • based on InsightFace in-swapper
  • PhotoMaker
    • for SD-XL only
    • new model from TenencentARC using similar concept as IPAdapter, but with different implementation and
      allowing full concept swaps between input images and generated images using trigger words
    • note: trigger word must match exactly one term in prompt for model to work
  • InstantID
    • for SD-XL only
    • based on custom trained ip-adapter and controlnet combined concepts
    • note: controlnet appears to be heavily watermarked
  • enable use via api, thanks @Trojaner
• IPAdapter
  • additional models for SD15 and SD-XL, to use simply select from Scripts:
    SD15: Base, Base ViT-G, Light, Plus, Plus Face, Full Face
    SDXL: Base SXDL, Base ViT-H SXDL, Plus ViT-H SXDL, Plus Face ViT-H SXDL
  • enable use via api, thanks @Trojaner
• Segmind SegMoE
  • initial support for reference models
    download&load via network -> models -> reference -> SegMoE SD 4x2 (3.7GB), SegMoE XL 2x1 (10GB), SegMoE XL 4x2
  • note: since segmoe is basically sequential mix of unets from multiple models, it can get large
    SD 4x2 is ~4GB, XL 2x1 is ~10GB and XL 4x2 is 18GB
  • supports lora, thanks @AI-Casanova
  • support for create and load custom mixes will be added in the future
• Mixture Tiling
  • uses multiple prompts to guide different parts of the grid during diffusion process
  • can be used ot create complex scenes with multiple subjects
  • simply select from scripts
• Self-attention guidance
  • simply select scale in advanced menu
  • can drastically improve image coherence as well as reduce artifacts
  • note: only compatible with some schedulers
• FreeInit for AnimateDiff
  • greatly improves temporal consistency of generated outputs
  • all options are available in animateddiff script
• SalesForce BlipDiffusion
  • model can be used to place subject in a different context
  • requires input image
  • last word in prompt and negative prompt will be used as source and target subjects
  • sampler must be set to default before loading the model
• InstaFlow
  • another take on super-fast image generation in a single step
  • set sampler:default, steps:1, cfg-scale:0
  • load from networks -> models -> reference
• Improvements
  • ui
    • check version and update SD.Next via UI
      simply go to: settings -> update
    • globally configurable font size
      will dynamically rescale ui depending on settings -> user interface
    • built-in themes can be changed on-the-fly
      this does not work with gradio-default themes as css is created by gradio itself
    • two new themes: simple-dark and simple-light
    • modularized blip/booru interrogate
      now appears as toolbuttons on image/gallery output
    • faster browser page load
    • update hints, thanks @brknsoul
    • cleanup settings
  • server
    • all move/offload options are disable by default for optimal performance
      enable manually if low on vram
  • server startup: performance
    • reduced module imports
      ldm support is now only loaded when running in backend=original
    • faster extension load
    • faster json parsing
    • faster lora indexing
    • lazy load optional imports
    • batch embedding load, thanks @midcoastal and @AI-Casanova
      10x+ faster embeddings load for large number of embeddings, now works for 1000+ embeddings
    • file and folder list caching, thanks @midcoastal
      if you have a lot of files and and/or are using slower or non-local storage, this speeds up file access a lot
    • add SD_INSTALL_DEBUG env variable to trace all git and pip operations
  • extra networks
    • 4x faster civitai metadata and previews lookup
    • better display and selection of tags & trigger words
      if hashes are calculated, trigger words will only be displayed for actual model version
    • better matching of previews
    • better search, including searching for multiple keywords or using full regex
      see wiki page for more details on syntax
      thanks @NetroScript
    • reduce html overhead
  • model compression, thanks @Disty0
    • using built-in NNCF model compression, you can reduce the size of your models significantly
      example: up to 3.4GB of VRAM saved for SD-XL model!
    • see wiki for details
  • embeddings
    you can now use sd 1.5 embeddings with your sd-xl models!, thanks @AI-Casanova
    conversion is done on-the-fly, is completely transparent and result is an approximation of embedding
    to enable: settings->extra networks->auto-convert embeddings
  • offline deployment: allow deployment without git clone
    for example, you can now deploy a zip of the sdnext folder
  • latent upscale: updated latent upscalers (some are new)
    nearest, nearest-exact, area, bilinear, bicubic, bilinear-antialias, bicubic-antialias
  • scheduler: added SA Solver
  • model load to gpu
    new option in settings->diffusers allowing models to be loaded directly to GPU while keeping RAM free
    this option is not compatible with any kind of model offloading as model is expected to stay in GPU
    additionally, all model-moves can now be traced with env variable SD_MOVE_DEBUG
  • xyz grid
    • range control
      example: 5.0-6.0:3 will generate 3 images with values 5.0,5.5,6.0
      example: 10-20:4 will generate 4 images with values 10,13,16,20
    • continue on error
      now you can use xyz grid with different params and test which ones work and which dont
    • correct font scaling, thanks @nCoderGit
  • hypertile
    • enable vae tiling
    • add autodetect optimial value
      set tile size to 0 to use autodetected value
  • cli
    • sdapi.py allow manual api invoke
      example: python cli/sdapi.py /sdapi/v1/sd-models
    • image-exif.py improve metadata parsing
    • install-sf helper script to automatically find best available stable-fast package for the platform
  • memory: add ram usage monitoring in addition to gpu memory usage monitoring
  • vae: enable taesd batch decode
    enable/disable with settings -> diffusers > vae slicing
• compile
  • new option: fused projections
    pretty much free 5% performance boost for compatible models
    enable in settings -> compute settings
  • new option: dynamic quantization (experimental)
    reduces memory usage and increases performance
    enable in settings -> compute settings
    best used together with torch compile: inductor
    this feature is highly experimental and will evolve over time
    requires nightly versions of torch and torchao

    pip install -U --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu121
    pip install -U git+https://github.com/pytorch-labs/ao

  • new option: compile text encoder (experimental)
• correction
  • new section in generate, allows for image corrections during generataion directly in latent space
  • adds brightness, sharpness and color controls, thanks @AI-Casanova
  • adds color grading controls, thanks @AI-Casanova
  • replaces old hdr section
• IPEX, thanks @Disty0
  • see wiki for details
  • rewrite ipex hijacks without CondFunc
    improves compatibilty and performance
    fixes random memory leaks
  • out of the box support for Intel Data Center GPU Max Series
  • remove IPEX / Torch 2.0 specific hijacks
  • add IPEX_SDPA_SLICE_TRIGGER_RATE, IPEX_ATTENTION_SLICE_RATE and IPEX_FORCE_ATTENTION_SLICE env variables
  • disable 1024x1024 workaround if the GPU supports 64 bit
  • fix lock-ups at very high resolutions
• OpenVINO, thanks @Disty0
  • see wiki for details
  • quantization support with NNCF
    run 8 bit directly without autocast
    enable OpenVINO Quantize Models with NNCF from Compute Settings
  • 4-bit support with NNCF
    enable Compress Model weights with NNCF from Compute Settings and set a 4-bit NNCF mode
    select both CPU and GPU from the device selection if you want to use 4-bit or 8-bit modes on GPU
  • experimental support for Text Encoder compiling
    OpenVINO is faster than IPEX now
  • update to OpenVINO 2023.3.0
  • add device selection to Compute Settings
    selecting multiple devices will use HETERO device
  • remove OPENVINO_TORCH_BACKEND_DEVICE env variable
  • reduce system memory usage after compile
  • fix cache loading with multiple models
• Olive support, thanks @lshqqytiger
  • fully merged in in wiki, see wiki for details
  • as a highlight, 4-5 it/s using DirectML on AMD GPU translates to 23-25 it/s using ONNX/Olive!
• fixes
  • civitai model download: enable downloads of embeddings
  • ipadapter: allow changing of model/image on-the-fly
  • ipadapter: fix fallback of cross-attention on unload
  • rebasin iterations, thanks @AI-Casanova
  • prompt scheduler, thanks @AI-Casanova
  • python: fix python 3.9 compatibility
  • sdxl: fix positive prompt embeds
  • img2img: clip and blip interrogate
  • img2img: sampler selection offset
  • img2img: support variable aspect ratio without explicit resize
  • cli: add simple-upscale.py script
  • cli: fix cmd args parsing
  • cli: add run-benchmark.py script
  • api: add /sdapi/v1/version endpoint
  • api: add /sdapi/v1/platform endpoint
  • api: return current image in progress api if requested
  • api: sanitize response object
  • api: cleanup error logging
  • api: fix api-only errors
  • api: fix image to base64
  • api: fix upscale
  • refiner: fix use of sd15 model as refiners in second pass
  • refiner: enable none as option in xyz grid
  • sampler: add sampler options info to metadata
  • sampler: guard against invalid sampler index
  • sampler: add img2img_extra_noise option
  • config: reset default cfg scale to 6.0
  • hdr: fix math, thanks @AI-Casanova
  • processing: correct display metadata
  • processing: fix batch file names
  • live preview: fix when using bfloat16
  • live preview: add thread locking
  • upscale: fix ldsr
  • huggingface: handle fallback model variant on load
  • reference: fix links to models and use safetensors where possible
  • model merge: unbalanced models where not all keys are present, thanks @AI-Casanova
  • better sdxl model detection
  • global crlf->lf switch
  • model type switch if there is loaded submodels
  • cleanup samplers use of compute devices, thanks @Disty0
• other
  • extensions sd-webui-controlnet is locked to commit ecd33eb due to breaking changes
  • extension stable-diffusion-webui-images-browser is locked to commit 27fe4a7 due to breaking changes
  • updated core requirements
  • fully dynamic pipelines
    pipeline switch is now done on-the-fly and does not require manual initialization of individual components
    this allows for quick implementation of new pipelines
    see modules/sd_models.py:switch_pipe for details
  • major internal ui module refactoring
    this may cause compatibility issues if an extension is doing a direct import from ui.py
    in which case, report it so we can add a compatibility layer
  • major public api refactoring
    this may cause compatibility issues if an extension is doing a direct import from api.py or models.py
    in which case, report it so we can add a compatibility layer

[vladmandic]

Update for 2024-02-22

Only 3 weeks since last release, but here's another feature-packed one!
This time release schedule was shorter as we wanted to get some of the fixes out faster.

Highlights

• IP-Adapters & FaceID: multi-adapter and multi-image suport
• New optimization engines: DeepCache, ZLUDA and Dynamic Attention Slicing
• New built-in pipelines: Differential diffusion and Regional prompting
• Big updates to: Outpainting (noised-edge-extend), Clip-skip (interpolate with non-integrer values!), CFG end (prevent overburn on high CFG scales), Control module masking functionality
• All reported issues since the last release are addressed and included in this release

Further details:

• For basic instructions, see README
• For more details on all new features see full CHANGELOG
• For documentation, see WiKi
• Discord server

Full ChangeLog for 2024-02-22

• Improvements:
  • IP Adapter major refactor
    • support for multiple input images per each ip adapter
    • support for multiple concurrent ip adapters
      note: you cannot mix & match ip adapters that use different CLiP models, for example Base and Base ViT-G
    • add adapter start/end to settings, thanks @AI-Casanova
      having adapter start late can help with better control over composition and prompt adherence
      having adapter end early can help with overal quality and performance
    • unified interface in txt2img, img2img and control
    • enhanced xyz grid support
  • FaceID now also works with multiple input images!
  • Differential diffusion
    img2img generation where you control strength of each pixel or image area
    can be used with manually created masks or with auto-generated depth-maps
    uses general denoising strength value
    simply enable from img2img -> scripts -> differential diffusion
    note: supports sd15 and sdxl models
  • Regional prompting as a built-in solution
    usage is same as original implementation from @hako-mikan
    click on title to open docs and see examples of full syntax on how to use it
    simply enable from scripts -> regional prompting
    note: supports sd15 models only
  • DeepCache model acceleration
    it can produce massive speedups (2x-5x) with no overhead, but with some loss of quality
    settings -> compute -> model compile -> deep-cache and settings -> compute -> model compile -> cache interval
  • ZLUDA experimental support, thanks @lshqqytiger
    • ZLUDA is CUDA wrapper that can be used for GPUs without native support
    • best use case is AMD GPUs on Windows, see wiki for details
  • Outpaint control outpaint now uses new alghorithm: noised-edge-extend
    new method allows for much larger outpaint areas in a single pass, even outpaint 512->1024 works well
    note that denoise strength should be increased for larger the outpaint areas, for example outpainting 512->1024 works well with denoise 0.75
    outpaint can run in img2img mode (default) and inpaint mode where original image is masked (if inpaint masked only is selected)
  • Clip-skip reworked completely, thanks @AI-Casanova & @Disty0
    now clip-skip range is 0-12 where previously lowest value was 1 (default is still 1)
    values can also be decimal to interpolate between different layers, for example clip-skip: 1.5, thanks @AI-Casanova
  • CFG End new param to control image generation guidance, thanks @AI-Casanova
    sometimes you want strong control over composition, but you want it to stop at some point
    for example, when used with ip-adapters or controlnet, high cfg scale can overpower the guided image
  • Control
    • when performing inpainting, you can specify processing resolution using size->mask
    • units now have extra option to re-use current preview image as processor input
  • Cross-attention refactored cross-attention methods, thanks @Disty0
    • for backend:original, its unchanged: SDP, xFormers, Doggettxs, InvokeAI, Sub-quadratic, Split attention
    • for backend:diffuers, list is now: SDP, xFormers, Batch matrix-matrix, Split attention, Dynamic Attention BMM, Dynamic Attention SDP
      note: you may need to update your settings! Attention Slicing is renamed to Split attention
    • for ROCm, updated default cross-attention to Scaled Dot Product
  • Dynamic Attention Slicing, thanks @Disty0
    • dynamically slices attention queries in order to keep them under the slice rate
      slicing gets only triggered if the query size is larger than the slice rate to gain performance
      Dynamic Attention Slicing BMM uses Batch matrix-matrix
      Dynamic Attention Slicing SDP uses Scaled Dot Product
    • settings -> compute settings -> attention -> dynamic attention slicing
  • ONNX:
    • allow specify onnx default provider and cpu fallback
      settings -> diffusers
    • allow manual install of specific onnx flavor
      settings -> onnx
    • better handling of fp16 models/vae, thanks @lshqqytiger
  • OpenVINO update to torch 2.2.0, thanks @Disty0
  • HyperTile additional options thanks @Disty0
    • add swap size option
    • add use only for hires pass option
  • add --theme cli param to force theme on startup
  • add --allow-paths cli param to add additional paths that are allowed to be accessed via web, thanks @OuticNZ
• Wiki:
  • added benchmark notes for IPEX, OpenVINO and Olive
  • added ZLUDA wiki page
• Internal
  • update dependencies
  • refactor txt2img/img2img api
  • enhanced theme loader
  • add additional debug env variables
  • enhanced sdp cross-optimization control
    see settings -> compute settings
  • experimental support for python 3.12
• Fixes:
  • add variation seed to diffusers txt2img, thanks @AI-Casanova
  • add cmd param --skip-env to skip setting of environment parameters during sdnext load
  • handle extensions that install conflicting versions of packages
    onnxruntime, opencv2-python
  • installer refresh package cache on any install
  • fix embeddings registration on server startup, thanks @AI-Casanova
  • ipex handle dependencies, thanks @Disty0
  • insightface handle dependencies
  • img2img mask blur and padding
  • xyz grid handle ip adapter name and scale
  • lazy loading of image may prevent metadata from being loaded on time
  • allow startup without valid models folder
  • fix interrogate api endpoint
  • control fix resize causing runtime errors
  • control fix processor override image after processor change
  • control fix display grid with batch
  • control restore pipeline before running scripts/extensions
  • handle pipelines that return dict instead of object
  • lora use strict name matching if preferred option is by-filename
  • fix inpaint mask only for diffusers
  • fix vae dtype mismatch, thanks @Disty0
  • fix controlnet inpaint mask
  • fix theme list refresh
  • fix extensions update information in ui
  • fix taesd with bfloat16
  • fix model merge manual merge settings, thanks @AI-Casanova
  • fix gradio instant update issues for textboxes in quicksettings
  • fix rembg missing dependency
  • bind controlnet extension to last known working commit, thanks @Aptronymist
  • prompts-from-file fix resizable prompt area

[vladmandic]

Update for 2024-03-19

Highlights 2024-03-19

New models:

• Stable Cascade Full and Lite
• Playground v2.5
• KOALA 700M
• Stable Video Diffusion XT 1.1
• VGen

New pipelines and features:

• Img2img using LEdit++, context aware method with image analysis and positive/negative prompt handling
• Trajectory Consistency Distillation TCD for processing in even less steps
• Visual Query & Answer using moondream2 as an addition to standard interrogate methods
• Face-HiRes: simple built-in detailer for face refinements
• Even simpler outpaint: when resizing image, simply pick outpaint method and if image has different aspect ratio, blank areas will be outpainted!
• UI aspect-ratio controls and other UI improvements
• User controllable invisibile and visible watermarking
• Native composable LoRA

What else?

• Reference models: Networks -> Models -> Reference: All reference models now come with recommended settings that can be auto-applied if desired
• Styles: Not just for prompts! Styles can apply generate parameters as templates and can be used to apply wildcards to prompts
  improvements, Additional API endpoints
• Given the high interest in ZLUDA engine introduced in last release we've updated much more flexible/automatic install procedure (see wiki for details)
• Plus Additional Improvements such as: Smooth tiling, Refine/HiRes workflow improvements, Control workflow

Further details:

• For basic instructions, see README
• For more details on all new features see full CHANGELOG
• For documentation, see WiKi
• Discord server

Full Changelog 2024-03-19

• Stable Cascade Full and Lite
  • large multi-stage high-quality model from warp-ai/wuerstchen team and released by stabilityai
  • download using networks -> reference
  • see wiki for details
• Playground v2.5
  • new model version from Playground: based on SDXL, but with some cool new concepts
  • download using networks -> reference
  • set sampler to DPM++ 2M EDM or Euler EDM
• KOALA 700M
  • another very fast & light sdxl model where original unet was compressed and distilled to 54% of original size
  • download using networks -> reference
  • note to download fp16 variant (recommended), set settings -> diffusers -> preferred model variant
• LEdit++
  • context aware img2img method with image analysis and positive/negative prompt handling
  • enable via img2img -> scripts -> ledit
  • uses following params from standard img2img: cfg scale (recommended ~3), steps (recommended ~50), denoise strength (recommended ~0.7)
  • can use postive and/or negative prompt to guide editing process
    • positive prompt: what to enhance, strength and threshold for auto-masking
    • negative prompt: what to remove, strength and threshold for auto-masking
  • note: not compatible with model offloading
• Second Pass / Refine
  • independent upscale and hires options: run hires without upscale or upscale without hires or both
  • upscale can now run 0.1-8.0 scale and will also run if enabled at 1.0 to allow for upscalers that simply improve image quality
  • update ui section to reflect changes
  • note: behavior using backend:original is unchanged for backwards compatibilty
• Visual Query visual query & answer in process tab
  • go to process -> visual query
  • ask your questions, e.g. "describe the image", "what is behind the subject", "what are predominant colors of the image?"
  • primary model is moondream2, a tiny 1.86B vision language model
    note: its still 3.7GB in size, so not really tiny
  • additional support for multiple variations of several base models: GIT, BLIP, ViLT, PIX, sizes range from 0.3 to 1.7GB
• Video
  • Image2Video
    • new module for creating videos from images
    • simply enable from img2img -> scripts -> image2video
    • model is auto-downloaded on first use
    • based on VGen
  • Stable Video Diffusion
    • updated with SVD 1.0, SVD XT 1.0 and SVD XT 1.1
    • models are auto-downloaded on first use
    • simply enable from img2img -> scripts -> stable video diffusion
    • for svd 1.0, use frames=~14, for xt models use frames=~25
• Composable LoRA, thanks @AI-Casanova
  • control lora strength for each step
    for example: <xxx:0.1@0,0.9@1> means strength=0.1 for step at 0% and intepolate towards strength=0.9 for step at 100%
  • note: this is a very experimental feature and may not work as expected
• Control
  • added refiner/hires workflows
  • added resize methods to before/after/mask: fixed, crop, fill
• Styles: styles are not just for prompts!
  • new styles editor: networks -> styles -> edit
  • styles can apply generate parameters, for example to have a style that enables and configures hires:
    parameters=enable_hr: True, hr_scale: 2, hr_upscaler: Latent Bilinear antialias, hr_sampler_name: DEIS, hr_second_pass_steps: 20, denoising_strength: 0.5
  • styles can apply wildcards to prompts, for example:
    wildcards=movie=mad max, dune, star wars, star trek; intricate=realistic, color sketch, pencil sketch, intricate
  • as usual, you can apply any number of styles so you can choose which settings are applied and in which order and which wildcards are used
• UI
  • aspect-ratio* add selector and lock to width/height control
    allowed aspect ration can be configured via settings -> user interface
  • interrogate tab is now merged into process tab
  • image viewer now displays image metadata
  • themes improve on-the-fly switching
  • log monitor flag server warnings/errors and overall improve display
  • control separate processor settings from unit settings
• Face HiRes
  • new face restore option, works similar to well-known adetailer by running an inpaint on detected faces but with just a checkbox to enable/disable
  • set as default face restorer in settings -> postprocessing
  • disabled by default, to enable simply check face restore in your generate advanced settings
  • strength, steps and sampler are set using by hires section in refine menu
  • strength can be overriden in settings -> postprocessing
  • will use secondary prompt and secondary negative prompt if present in refine
• Watermarking
  • SD.Next disables all known watermarks in models, but does allow user to set custom watermark
  • see settings -> image options -> watermarking
  • invisible watermark: using steganogephy
  • image watermark: overlaid on top of image
• Reference models
  • additional reference models available for single-click download & run:
    Stable Cascade, Stable Cascade lite, Stable Video Diffusion XT 1.1
  • reference models will now download fp16 variation by default
  • reference models will print recommended settings to log if present
  • new setting in extra network: use reference values when available
    disabled by default, if enabled will force use of reference settings for models that have them
• Samplers
  • TCD: Trajectory Consistency Distillation
    new sampler that produces consistent results in a very low number of steps (comparable to LCM but without reliance on LoRA)
    for best results, use with TCD LoRA: https://huggingface.co/h1t/TCD-SDXL-LoRA
  • DPM++ 2M EDM and Euler EDM
    EDM is a new solver algorithm currently available for DPM++2M and Euler samplers
    Note that using EDM samplers with non-EDM optimized models will provide just noise and vice-versa
• Improvements
  • FaceID extend support for LoRA, HyperTile and FreeU, thanks @Trojaner
  • Tiling now extends to both Unet and VAE producing smoother outputs, thanks @AI-Casanova
  • new setting in image options: include mask in output
  • improved params parsing from from prompt string and styles
  • default theme updates and additional built-in theme black-gray
  • support models with their own YAML model config files
  • support models with their own JSON per-component config files, for example: playground-v2.5_vae.config
  • prompt can have comments enclosed with /* and */
    comments are extracted from prompt and added to image metadata
• ROCm
  • add ROCm 6.0 nightly option to installer, thanks @jicka
  • add flash attention support for rdna3, thanks @Disty0
    install flash_attn package for rdna3 manually and enable flash attention from compute settings
    to install flash_attn, activate the venv and run pip install -U git+https://github.com/ROCm/flash-attention@howiejay/navi_support
• IPEX
  • disabled IPEX Optimize by default
• API
  • add preprocessor api endpoints
    GET:/sdapi/v1/preprocessors, POST:/sdapi/v1/preprocess, sample script:cli/simple-preprocess.py
  • add masking api endpoints
    GET:/sdapi/v1/masking, POST:/sdapi/v1/mask, sample script:cli/simple-mask.py
• Internal
  • improved vram efficiency for model compile, thanks @Disty0
  • stable-fast compatibility with torch 2.2.1
  • remove obsolete textual inversion training code
  • remove obsolete hypernetworks training code
• Refiner validated workflows:
  • Fully functional: SD15 + SD15, SDXL + SDXL, SDXL + SDXL-R
  • Functional, but result is not as good: SD15 + SDXL, SDXL + SD15, SD15 + SDXL-R
• SDXL Lightning models just-work, just makes sure to set CFG Scale to 0
  and choose a best-suited sampler, it may not be the one you're used to (e.g. maybe even basic Euler)
• Fixes
  • improve model cpu offload compatibility
  • improve model sequential offload compatibility
  • improve bfloat16 compatibility
  • improve xformers installer to match cuda version and install triton
  • fix extra networks refresh
  • fix sdp memory attention in backend original
  • fix autodetect sd21 models
  • fix api info endpoint
  • fix sampler eta in xyz grid, thanks @AI-Casanova
  • fix requires_aesthetics_score errors
  • fix t2i-canny
  • fix differenital diffusion for manual mask, thanks @23pennies
  • fix ipadapter apply/unapply on batch runs
  • fix control with multiple units and override images
  • fix control with hires
  • fix control-lllite
  • fix font fallback, thanks @NetroScript
  • update civitai downloader to handler new metadata
  • improve control error handling
  • use default model variant if specified variant doesnt exist
  • use diffusers lora load override for lcm/tcd/turbo loras
  • exception handler around vram memory stats gather
  • improve ZLUDA installer with --use-zluda cli param, thanks @lshqqytiger

[vladmandic]

Update for 2024-05-28

Highlights for 2024-05-28

New SD.Next release has been baking in dev for a longer than usual, but changes are massive - about 350 commits for core and 300 for UI...

Starting with the new UI - yup, this version ships with a preview of the new ModernUI
For details on how to enable and use it, see Home and WiKi
ModernUI is still in early development and not all features are available yet, please report issues and feedback
Thanks to @BinaryQuantumSoul for his hard work on this project!

What else?

New built-in features

• PWA SD.Next is now installable as a web-app
• Gallery: extremely fast built-in gallery viewer
  List, preview, search through all your images and videos!
• HiDiffusion allows generating very-high resolution images out-of-the-box using standard models
• Perturbed-Attention Guidance (PAG) enhances sample quality in addition to standard CFG scale
• LayerDiffuse simply create transparent (foreground-only) images
• IP adapter masking allows to use multiple input images for each segment of the input image
• IP adapter InstantStyle implementation
• Token Downsampling (ToDo) provides significant speedups with minimal-to-none quality loss
• Samplers optimizations that allow normal samplers to complete work in 1/3 of the steps!
  Yup, even popular DPM++2M can now run in 10 steps with quality equaling 30 steps using AYS presets
• Native wildcards support
• Improved built-in Face HiRes
• Better outpainting
• And much more...
  For details of above features and full list, see Changelog

New models

While still waiting for Stable Diffusion 3.0, there have been some significant models released in the meantime:

• PixArt-Σ, high end diffusion transformer model (DiT) capable of directly generating images at 4K resolution
• SDXS, extremely fast 1-step generation consistency model
• Hyper-SD, 1-step, 2-step, 4-step and 8-step optimized models

Note
SD.Next is no longer marked as a fork of A1111 and github project has been fully detached
Given huge number of changes with +3443/-3342 commits diff (at the time of fork detach) over the past year,
a completely different backend/engine and a change of focus, it is time to give credit to original author, and move on!

Full ChangeLog for 2024-05-28

• Features:
  • ModernUI preview of the new ModernUI
  • PWA SD.Next is now installable as a web-app and includes verified manifest
  • **Gallery
  • Gallery: list, preview, search through all your images and videos!
    Implemented as infinite-scroll with client-side-caching and lazy-loading while being fully async and non-blocking
    Search or sort by path, name, size, width, height, mtime or any image metadata item, also with extended syntax like width > 1000
    Settings: optional additional user-defined folders, thumbnails in fixed or variable aspect-ratio
  • HiDiffusion:
    Generate high-resolution images using your standard models without duplicates/distorsions AND improved performance
    For example, SD15 can now go up to 2024x2048 and SDXL up to 4k natively
    Simply enable checkbox in advanced menu and set desired resolution
    Additional settings are available in settings -> inference settings -> hidiffusion
    And can also be set and used via xyz grid
    Note: HiDiffusion resolution sensitive, so if you get error, set resolution to be multiples of 128
  • Perturbed-Attention Guidance
    PAG enhances sample quality by utilizing self-attention in formation of latent in addition to standard CFG scale
    Simply set advanced -> attention guidance and advanced -> adaptive scaling
    Additional options are available in settings -> inference settings -> pag
    Note: PAG has replaced SAG as attention guidance method in SD.Next
  • LayerDiffuse
    Create transparent images with foreground-only being generated
    Simply select from scripts -> apply to current model
    All necessary files will be auto-downloaded on first use
  • IP Adapter Masking:
    Powerful method of using masking with ip-adapters
    When combined with multiple ip-adapters, it allows for different inputs guidance for each segment of the input image
    Hint: to create masks, you can use manually created masks or control->mask module with auto-segment to create masks and later upload them
  • IP Adapter advanced layer configuration:
    Allows for more control over how each layer of ip-adapter is applied, requires a valid dict to be passed as input
    See InstantStyle for details
  • OneDiff: new optimization/compile engine, thanks @aifartist
    As with all other compile engines, enable via settings -> compute settings -> compile
  • ToDo Token Downsampling for Efficient Generation of High-Resolution Images
    Newer alternative method to ToMe that can provide speed-up with minimal quality loss
    Enable in settings -> inference settings -> token merging
    Also available in XYZ grid
  • Outpaint:
    New method of outpainting that uses a combination of auto-masking and edge generation to create seamless transitions between original and generated image
    Use on control tab:
    • input -> denoising strength: 0.5 or higher
    • select image -> outpaint -> expand edges or zoom out to desired size
    • size -> mode: outpaint, method: nearest
    • mask -> inpaint masked only (if you want to keep original image)
  • Wildcards:
    • native support of standard file-based wildcards in prompt
    • enabled by default, can be disabled in settings -> extra networks if you want to use 3rd party extension
    • wildcards folder is set in settings -> system paths and can be flat-file list or complex folder structure
    • matches strings "__*__" in positive and negative prompts
    • supports filename and path-based wildcards
    • supports nested wildcards (wildcard can refer to another wildcard, etc.)
    • supports wildcards files in one-choice per line or multiple choices per line separated by | format
    • note: this is in addition to previously released style-based wildcards
• Models:
  • Load UNET: ability to override/load external UNET to a selected model
    Works similar to how VAE is selected and loaded: Set UNet folder and UNet model in settings
    Can be replaced on-the-fly, not just during initial model load
    Enables usage of fine-tunes such as DPO-SD15 or DPO-SDXL
    Note: if there is a JSON file with the same name as the model it will be used as Unet config, otherwise Unet config from currently loaded model will be used
  • PixArt-Σ
    pixart-Σ is a high end diffusion Transformer model (DiT) with a T5 encoder/decoder capable of directly generating images at 4K resolution
    to use, simply select from networks -> models -> reference -> PixArt-Σ
    note: this is a very large model at ~22GB
    set parameters: sampler: Default
  • SDXS
    sdxs is an extremely fast 1-step generation consistency model that also uses TAESD as quick VAE out-of-the-box
    to use, simply select from networks -> models -> reference -> SDXS
    set parameters: sampler: CMSI, steps: 1, cfg_scale: 0.0
  • Hyper-SD
    sd15 and sdxl 1-step, 2-step, 4-step and 8-step optimized models using lora
    set parameters: sampler: TCD or LCM, steps: 1/2/4/8, cfg_scale: 0.0
• UI:
  • Faster UI load times
  • Theme types:
    Standard (built-in themes), Modern (experimental nextgen ui), None (used for Gradio and Huggingface 3rd party themes)
    Specifying a theme type updates list of available themes
    For example, Gradio themes will not appear as available if theme type is set to Standard
  • Redesign of base txt2img interface
  • Minor tweaks to styles: refresh/apply/save
  • See details in WiKi
• API:
  • Add API endpoint /sdapi/v1/control and CLI util cli/simple-control.py
    (in addition to previously added /sdapi/v1/preprocessors and /sdapi/v1/masking)
    example:

    simple-control.py --prompt 'woman in the city' --sampler UniPC --steps 20
    --input ~/generative/Samples/cutie-512.png --output /tmp/test.png --processed /tmp/proc.png
    --control 'Canny:Canny FP16:0.7, OpenPose:OpenPose FP16:0.8' --type controlnet
    --ipadapter 'Plus:~/generative/Samples/cutie-512.png:0.5'

  • Add API endpoint /sdapi/v1/vqa and CLI util cli/simple-vqa.py
• Changes:
  • Due to change in Diffusers model loading
    initial model load will now fetch config files required for the model
    from the Huggingface site instead of using predefined YAML files
  • Removed built-in extensions: ControlNet and Image-Browser
    as both image-browser and controlnet have native built-in equivalents
    both can still be installed by user if desired
  • Different defaults depending on available GPU, thanks @Disty0
    • 4GB and below: lowvram
    • 8GB and below: medvram
    • Cross-attention: Dynamic Attention SDP with medvram or lowvram, otherwise SDP
    • VAE Tiling enabled with medvram and lowvram
    • Disable Extract EMA by default
    • Disable forced VAE Slicing for lowvram
  • Upscaler compile disabled by default with OpenVINO backend
  • Hypernetwork support disabled by default, can be enabled in settings
• Improvements:
  • Faster server startup
  • Styles apply wildcards to params
  • Face HiRes fully configurable and higher quality when using high-resolution models
  • Extra networks persistent sort order in settings
  • Add option to make batch generations use fully random seed vs sequential
  • Make metadata in full screen viewer optional
  • Add VAE civitai scan metadata/preview
  • More efficient in-browser callbacks
  • Updated all system requirements
  • UI log monitor will auto-reconnect to server on server restart
  • UI styles includes indicator for active styles
  • UI reduce load on browser
  • Secondary sampler add option "same as primary"
  • Change attention mechanism on-the-fly without model reload, thanks @Disty0
  • Update stable-fast with support for torch 2.2.2 and 2.3.0, thanks @Aptronymist
  • Add torch cudaMallocAsync in compute options
    Can improve memory utilization on compatible GPUs (RTX and newer)
  • Torch dynamic profiling
    You can enable/disable full torch profiling in settings top menu on-the-fly
  • Prompt caching - if you use the same prompt multiple times, no need to re-parse and encode it
    Useful for batches as prompt processing is ~0.1sec on each pass
  • Enhance SD_PROMPT_DEBUG to show actual tokens used
  • Support controlnet manually downloads models in both standalone and diffusers format
    For standalone, simply copy safetensors file to models/control/controlnet folder
    For diffusers format, create folder with model name in models/control/controlnet/
    and copy model.json and diffusion_pytorch_model.safetensors to that folder
• Samplers
  • Add Euler SGM variation (e.g. SGM Uniform), optimized for SDXL-Lightning models
    note: you can use other samplers as well with SDXL-Lightning models
  • Add CMSI sampler, optimized for consistency models
  • Add option timestep spacing to sampler settings and sampler section in main ui
    Note: changing timestep spacing changes behavior of sampler and can help to make any sampler turbo/lightning compatibile
  • Add option timesteps to manually set timesteps instead of relying on steps+spacing
    Additionally, presets from nVidias align-you-steps reasearch are provided
    Result is that perfectly aligned steps can drastically reduce number of steps needed!
    For example, AYS preset alows DPM++2M to run in ~10 steps with quality equallying ~30 steps!
• IPEX, thanks @Disty0
  • Update to IPEX 2.1.20 on Linux
    requires removing the venv folder to update properly
  • Removed 1024x1024 workaround
  • Disable ipexrun by default, set IPEXRUN=True if you want to use ipexrun
• ROCm, thanks @Disty0
  • Add support for ROCm 6.1 nighthly builds
  • Switch to stable branch of PyTorch
  • Compatibility improvenments
  • Add MIGraphX torch compile engine
• ZLUDA, thanks @lshqqytiger
  • Rewrite ZLUDA installer
  • ZLUDA v3.8 updates: Runtime API support
  • Add --reinstall-zluda (to download the latest ZLUDA)
• Fixes:
  • Update requirements
  • Installer automatically handle detached git states
  • Prompt params parser
  • Allowing forcing LoRA loading method for some or all models
  • Image save without metadata
  • API generate save metadata
  • Face/InstantID faults
  • CivitAI update model info for all models
  • FP16/BF16 test on model load
  • Variation seed possible NaNs
  • Enumerate diffusers model with multiple variants
  • Diffusers skip non-models on enum
  • Face-HiRes compatibility with control modules
  • Face-HiRes avoid doule save in some scenarios
  • Loading safetensors embeddings
  • CSS fixes
  • Check if attention processor is compatible with model
  • SD Upscale when used with control module
  • Noise sampler seed, thanks @leppie
  • Control module with ADetailer and active ControlNet
  • Control module restore button full functionality
  • Control improved handling with multiple control units and different init images
  • Control add correct metadata to image
  • Time embeddings load part of model load
  • A1111 update OptionInfo properties
  • MOTD exception handling
  • Notifications not triggering
  • Prompt cropping on copy

[vladmandic]

Update for 2024-06-13

Highlights for 2024-06-13

First, yes, it's here and supported: StabilityAI Stable Diffusion 3 Medium
for details on how to download and use, see Wiki

What else?

A lot of work on state-of-the-art multi-lingual models with both Tenecent HunyuanDiT and MuLan
Plus tons of minor features such as optimized initial install experience, T-Gate and ResAdapter, additional ModernUI themes (both light and dark) and fixes since the last release which was only 2 weeks ago!

Full Changelog for 2024-06-13

New Models

• StabilityAI Stable Diffusion 3 Medium
  yup, supported!
  quote: "Stable Diffusion 3 Medium is a multimodal diffusion transformer (MMDiT) model that features improved performance in image quality, typography, complex prompt understanding, and resource-efficiency"
  sdnext also supports switching optional T5 text encoder on-the-fly as well as loading model from either diffusers repo or safetensors single-file
  for details, see Wiki
• Tenecent HunyuanDiT bilingual english/chinese diffusion transformer model
  note: this is a very large model at ~17GB, but can be used with less VRAM using model offloading
  simply select from networks -> models -> reference, model will be auto-downloaded on first use

New Functionality

• MuLan Multi-langunage prompts
  write your prompts forin ~110 auto-detected languages!
  compatible with SD15 and SDXL
  enable in scripts -> MuLan and set encoder to InternVL-14B-224px encoder
  note: right now this is more of a proof-of-concept before smaller and/or quantized models are released
  model will be auto-downloaded on first use: note its huge size of 27GB
  even executing it in FP16 will require ~16GB of VRAM for text encoder alone
  examples:
  • English: photo of a beautiful woman wearing a white bikini on a beach with a city skyline in the background
  • Croatian: fotografija lijepe žene u bijelom bikiniju na plaži s gradskim obzorom u pozadini
  • Italian: Foto di una bella donna che indossa un bikini bianco su una spiaggia con lo skyline di una città sullo sfondo
  • Spanish: Foto de una hermosa mujer con un bikini blanco en una playa con un horizonte de la ciudad en el fondo
  • German: Foto einer schönen Frau in einem weißen Bikini an einem Strand mit einer Skyline der Stadt im Hintergrund
  • Arabic: صورة لامرأة جميلة ترتدي بيكيني أبيض على شاطئ مع أفق المدينة في الخلفية
  • Japanese: 街のスカイラインを背景にビーチで白いビキニを着た美しい女性の写真
  • Chinese: 一个美丽的女人在海滩上穿着白色比基尼的照片, 背景是城市天际线
  • Korean: 도시의 스카이라인을 배경으로 해변에서 흰색 비키니를 입은 아름 다운 여성의 사진
• T-Gate Speed up generations by gating at which step cross-attention is no longer needed
  enable via scripts -> t-gate
  compatible with SD15
• PCM LoRAs allow for fast denoising using less steps with standard SD15 and SDXL models
  download from https://huggingface.co/Kijai/converted_pcm_loras_fp16/tree/main
• ByteDance ResAdapter resolution-free model adapter
  allows to use resolutions from 0.5 to 2.0 of original model resolution, compatible with SD15 and SDXL
  enable via scripts -> resadapter and select desired model
• Kohya HiRes Fix allows for higher resolution generation using standard SD15 models
  enable via scripts -> kohya-hires-fix
  note: alternative to regular hidiffusion method, but with different approach to scaling
• additional built-in 4 great custom trained ControlNet SDXL models from Xinsir: OpenPose, Canny, Scribble, AnimePainter
  thanks @lbeltrame
• add torch full deterministic mode
  enable in settings -> compute -> use deterministic mode
  typical differences are not large and its disabled by default as it does have some performance impact
• new sampler: Euler FlowMatch

Improvements

• additional modernui themes
• reintroduce prompt attention normalization, disabled by default, enable in settings -> execution
  this can drastically help with unbalanced prompts
• further work on improving python 3.12 functionality and remove experimental flag
  note: recommended version remains python 3.11 for all users except if you're using directml and then its python 3.10
• improved installer for initial installs
  initial install will do single-pass install of all required packages with correct versions
  subsequent runs will check package versions as necessary
• add env variable SD_PIP_DEBUG to write pip.log for all pip operations
  also improved installer logging
• add python version check for torch-directml
• do not install tensorflow by default
• improve metadata/infotext parser
  add cli/image-exif.py that can be used to view/extract metadata from images
• lower overhead on generate calls
• auto-synchronize modernui and core branches
• add option to pad prompt with zeros, thanks @Disty

Fixes

• cumulative fixes since the last release
• fix apply/unapply hidiffusion for sd15
• fix controlnet reference enabled check
• fix face-hires with control batch count
• install pynvml on-demand
• apply rollback-vae option to latest torch versions, thanks @Iaotle
• face hires skip if strength is 0
• restore all sampler configuration on sampler change

Update for 2024-06-02

• fix textual inversion loading
• fix gallery mtime display
• fix extra network scrollable area when using modernui
• fix control prompts list handling
• fix restore variation seed and strength
• fix negative prompt parsing from metadata
• fix stable cascade progress monitoring
• fix variation seed with hires pass
• fix loading models trained with onetrainer
• add variation seed info to metadata
• workaround for scale-by when using modernui
• lock torch-directml version
• improve xformers installer
• improve ultralytics installer (face-hires)
• improve triton installer (compile)
• improve insightface installer (faceip)
• improve mim installer (dwpose)
• add dpm++ 1s and dpm++ 3m aliases for dpm++ 2m scheduler with different orders

[vladmandic]

SD.Next release 2024-06-23

Highlights

Following zero-day SD3 release, a 10 days later here's a refresh with 10+ improvements!
Including full prompt attention, support for compressed weights, additional text-encoder quantization modes.

But there's more than SD3:

• Support for quantized T5 text encoder FP16/FP8/FP4/INT8 in all models that use T5: SD3, PixArt-Σ, etc.
• Support for PixArt-Sigma in small/medium/large variants
• Support for HunyuanDiT 1.1
• Additional NNCF weights compression support: SD3, PixArt, ControlNet, Lora
• Integration of MS Florence VLM/VQA Base and Large models
• (finally) New release of Torch-DirectML
• Additional efficiencies for users with low VRAM GPUs
• Over 20 overall fixes

Model Improvements

• SD3: enable tiny-VAE (TAESD) preview and non-full quality mode
• SD3: enable base LoRA support
• SD3: add support for FP4 quantized T5 text encoder
  simply select in settings -> model -> text encoder
  note for SD3 with T5, set SD.Next to use FP16 precision, not BF16 precision
• SD3: add support for INT8 quantized T5 text encoder, thanks @Disty0!
• SD3: enable cpu-offloading for T5 text encoder, thanks @Disty0!
• SD3: simplified loading of model in single-file safetensors format
  model load can now be performed fully offline
• SD3: full support for prompt parsing and attention, thanks @AI-Casanova!
• SD3: ability to target different prompts to each of text-encoders, thanks @AI-Casanova!
  example: dog TE2: cat TE3: bird
• SD3: add support for sampler shift for Euler FlowMatch
  see settings -> samplers, also available as param in xyz grid
  higher shift means model will spend more time on structure and less on details
• SD3: add support for selecting T5 text encoder variant in XYZ grid
• Pixart-Σ: Add small (512px) and large (2k) variations, in addition to existing medium (1k)
• Pixart-Σ: Add support for 4/8bit quantized t5 text encoder
  note by default pixart-Σ uses full fp16 t5 encoder with large memory footprint
  simply select in settings -> model -> text encoder before or after model load
• HunyuanDiT: support for model version 1.1
• MS Florence: integration of Microsoft Florence VLM/VQA Base and Large models
  simply select in process -> visual query!

General Improvements

• support FP4 quantized T5 text encoder, in addtion to existing FP8 and FP16
• support for T5 text-encoder loader in all models that use T5
  example: load FP4 or FP8 quantized T5 text-encoder into PixArt Sigma!
• support for torch-directml 0.2.2, thanks @lshqqytiger!
  note: new directml is finally based on modern torch 2.3.1!
• xyz grid: add support for LoRA selector
• vae load: store original vae so it can be restored when set to none
• extra networks: info display now contains link to source url if model if its known
  works for civitai and huggingface models
• force gc for lowvram users and improve gc logging
• improved google.colab support
• css tweaks for standardui
• css tweaks for modernui
• additional torch gc checks, thanks @Disty0!

Improvements: NNCF, thanks @Disty0!

• SD3 and PixArt support
• moved the first compression step to CPU
• sequential cpu offload (lowvram) support
• Lora support without reloading the model
• ControlNet compression support

Fixes

• fix unsaturated outputs, force apply vae config on model load
• fix hidiffusion handling of non-square aspect ratios, thanks @ShenZhang-Shin!
• fix control second pass resize
• fix hunyuandit set attention processor
• fix civitai download without name
• fix compatibility with latest adetailer
• fix invalid sampler warning
• fix starting from non git repo
• fix control api negative prompt handling
• fix saving style without name provided
• fix t2i-color adapter
• fix sdxl "has been incorrectly initialized"
• fix api face-hires
• fix api ip-adapter
• fix memory exceptions with ROCm, thanks @Disty0!
• fix face-hires with lowvram, thanks @Disty0!
• fix pag incorrectly resetting pipeline
• cleanup image metadata
• restructure api examples: cli/api-*
• handle theme fallback when invalid theme is specified
• remove obsolete training code leftovers