SD.Next Release 2026-06-16

[vladmandic]

SD.Next Release 2026-06-16

What's New?

• Ideogram-4 released, Microsoft joins the game with Lens and Anima made it to release version
• SDNQ new quantization algorithm with even higher quality
• New Image Analysis feature and much improved Prompt Enhance capabilities which allow steering the model in real-time
• New workflows with ability to run Detailing as post-processing on existing images
• Updates to Kanvas, SD.Next Launcher and Enso UI!

And we have new Home page with heavily updated Docs and new Contributing & Development section in docs with info on pretty much any type of development or contribution related topics - do check it out!

Plus continued work on modernization of codebase: UI is now fully TypeScript based
And we have a new modular LoRA loader, new native Transformers loader and improved 3rd party finetunes support!

Note: This is a major update due to sheer size of the changes: over 400 commits!

Home | ChangeLog | Docs | Discord | Sponsor

Details for 2026-06-16

• Models
  • CircleStone Anima 1.0 in Base and Turbo (distilled) variants
    in both original precision and SDNQ-4bit quantiztion
  • Microsoft Lens in Standard, Base and Turbo (distilled) variant
    3.8B text-to-image DiT model with 12B GPT-OSS text-encoding and Flux2 VAE
    oh, that 12B encoder is MoE with 3.6B activated plus its prequantized using mxfp4
    note Lens comes with its own prompt-refiner, enable in settings -> model options (disabled by default)
    note original Lens implements only text-2-image, SD.Next adds image-2-image and inpaint workflows as well
  • Ideogram 4 open-weight 9.3B flow-matching single-stream DiT
    with dual-transformer architecture (9.5b) and qwen3 (8b) text encoder
    too many notes to add here, check out the dedicated Ideogram-4 wiki page for all details!
• Features
  • SDNQ new quantization algorithm: Hadamard Rotations
    much higher quality than base SDNQ, but runs slightly slower
    still faster than SVD and can be combined together with SVD for combined benefits
  • Captioning
    new feature: analyze existing images for prompt adherence
    tip: image analysis requires larger VLM model to produce quality output
    new api endpoint: /sdapi/v1/analyze
    cleanup list of predefined models, new models added and some old removed
    add support for prequantized models
    improved default values plus some new params like min length and custom args so you can pass anything to an llm model
    improved system prompts
  • Prompt Enhance tons of features
    cleanup list of predefined models, new models added and some old removed
    improved default values plus some new params like min length and custom args so you can pass anything to an llm model
    improved system prompts
    add support for prequantized models
    new processing engine! now you can steer the model as its generating
    add words to list and model will either steer away from them towards safe choices or you choose specific replacements for them
    for example: child:person, toy:airplane, dog:cat
    will do exactly as you'd expect, steer away from first word towards (optional) second word
    and it expands the functionality with customizable embedding similarity:
    for example, child can match kid, girl, boy
    and it expands the functionality with customizable semantic matching:
    for example, young ... will match before next word appears in the prompt and steer away from it towards desired choices
  • Detailer available as post-processing task for existing images
  • Masking updated interface and capabilities
    you can now also select mask type instead of focing alpha mask with all models
  • Samplers reorganized into clear sampler categories
  • Gallery add clear cache button to folder menu
  • Finetunes improved support for loading model finetunes
    this also includes detecting compatibility and falbacks
  • UV much updated --uv support for fast installs
    now also supports global uv if present in the system
  • Attention Dispatcher new attention backends dispatcher
    in settings -> attention Dispatcher
    allows to use pluggable kernels defines either in packages or in new kernels library
    see backends for list of available attention backends
    note compatibility matrix between torch backend, torch version and model specifics is relatively small at the moment
    note does not replace existing attention settings
  • Image metadata add wildcards (if used) info to image
    if wildcards or styles modify prompt, add original prompt to image metadata as template
  • Video metadata add processing info to video metadata as well, thanks @ryanmeador
  • Kanvas image resize is now two-way, you can resize in kanvas or in main ui
• Changes
  • all Guidance params are now set to -1 by default to allow using model defaults and avoid confusion with different model behaviour
    log will print default values used by model if not set by user
  • Shared components additional support for shared model components
    avoids unnecessary downloads and allows to share components between different models
    enabled by default, see settings -> text encoder -> use shared instance
  • restore params from image metadata will now prefer template field if present, otherwise use prompt field
    this allows to preserve original prompt in case of wildcards or styles modifying the prompt
  • HF download use XET by default
    see settings -> huggingface -> download method for options
  • Nunchaku consider DEV builds when auto-installing
  • Video save image thumbnail for generated video is now optional, thanks @ryanmeador
• Docs
  • new Contributing & Development home page
    includes pages on development setup, code structure, coding standards, ui development, themes, docs, hints and more!
• AI
  • Cognitive analysis and improvements to all AI prompts
  • Automated fixes using /check- skills
  • Automated syntax, spelling and readability improvements to /wiki pages
  • Multiple quality fixes based on CoPilot review
  • Multiple quality fixes based on Claude review, thanks @QualiaRain
• Internal
  • new native transformers loader!
    massive new codebase, but improves modularity and compatibility with different model architectures
  • new model finetunes loader!
    better compatibility for different finetunes and automatic detection of compatibility with base model
  • refactor shared components loader
    now takes into effect desired pre-quant precision and allows to share components between different models
  • update all model loading code to use consistent paths:
    diffusers_dir for image pipelines
    hfcache_dir for model components and auxiliary models
  • update torch==2.12 for CUDA, ROCm, IPEX
  • complete refactor of core JavaScript codebase to TypeScript!
  • complete refactor of modernui JavaScript codebase to TypeScript!
  • remove of /html and /javascript folders
  • add /ui folder for all ui-related code/css/assets
  • large refactor of lora native loader
  • improve kanvas typing
  • additional strong typing in core, thanks @awsr
  • full codespell coverage
    note this resulted in large one-time changeset
  • enhance automated testing
    pnpm test (uses --test) flag runs pipeline init checks
    pnpm compile (new) runs static python compile and import checks
• Fixes
  • hidream-o1 prequant loading
  • gradio initial hijack
  • SmolVLM captioning
  • gradio temp files guard against large image
  • diffusers patch custom pipelines for qk_norm
  • GHSA fixes, thanks @SSJCorpSec for reporting
  • custom vae loader
  • attention execution guard against cpu tensors
  • downloaded diffuser model use snapshot path
  • improve settings search
  • nunchaku z-image loader
  • ui server log monitor
  • kanvas enable toolbar on send-to action
  • hf download model card lookup
  • ltx video padding logic
  • prompt enhance custom model loader
  • styles loader exception handling
  • kanvas image change notification
  • reinstall force reinstall of transformers and diffusers
  • ipex torch install error, thanks @liutyi
  • taesd preview constant size with reduced layers
  • output path use correct base folder for initial folders
  • ltx prompt embeds move to device, thanks @ryanmeador
  • openpose processor
  • img2img api default sampler
  • sdnq default dynamic loss value
  • sdnq prequant save/load
  • samplers ui sigma methods
  • xpu generator on non-cpu
  • compel compatibility with transformers==5
  • gallery open folder
  • seedvr unload after upscale
  • tinyvae with anima
  • mixture-tiling fix for non-square images, thanks @QualiaRain
  • prompts-from-file fix metadata handling, thanks @QualiaRain
  • hypertile correct width/height assignment, thanks @QualiaRain
  • custom allowed-paths, thanks @QualiaRain
  • bias dtype, thanks @QualiaRain
  • ipadapter mask accumulation, thanks @QualiaRain
  • lora no-lora check, thanks @QualiaRain
  • control video processing, thanks @QualiaRain
  • freescale correct width/height assignment, thanks @QualiaRain
  • noise lerp inversion, thanks @QualiaRain
  • additional safety checks, thanks @QualiaRain
  • remote vae shadowing, thanks @QualiaRain