preview7 - 0.1.6

Added features

• CivitAi.com model downloader/browser

  Limitations

  • download history not saved
  • when the window is closed, the download will be aborted without user notification
  • aborted downloads part file (modelname.download) will not be deleted, need to be delete manually or it will be overwritten when a new download start with the same model file
  • no auto hash calculation after downloads
  • only locally hashed models will be marked in the CivitAi downloader as locally available model (if you not hashed a model file, and you will download the same file with same name, the old model file will be overwritten without notification)

• CivitAi api key for private model download (CivitAi docs). You can add your key in the Settings -> CivitAi tab

Modifications

• limit target image resolutions, only allowed resolutions with dividet by 256 (except for loading settings from old images)
• wxWidgets now statically linked into the binary

Some bug fixed in the linux versions. The deb files are created on Ubuntu 22. Now the gnome-secret / ksecretsservice etc.. is required to store the CivitAi api key.
The CPU and VGA versions are now in separated .deb files.

preview7 - 0.1.5

This preview release have some gui enhancement and inner data handling refactoring. This may cause unexpected behaviours. If you experience it, please report it in a new bug on the github's issues page.

Added features

• On windows and linux, dynamically load the library which is required / best to (currently available rocm, cuda, avx, avx2, avx512) the HW
• fully reworked build system, no more cmake just chalet
• on linux, the dependencies are depends from the OS, not vcpkg (supported wxwidgets 3.2 on linux)
• preview images are now on the bottom of the lists (queue job list & model info llist)
• some gui behaviour is just changed. When settings is open, the main ui will freezee and hide to disable unexpected behaviours on settings change
• added the following args to override the detected HW configurations:
  • "-cuda" to force to load cuda
  • "-rocm" to force to load rocm
  • "-avx"
  • "-avx2"
  • "-avx512"
  • "-allow-multiple" to allow multiple instances to run. Eg. if you wamt to start a cuda and an avx512 instance too:
    sd.gui.exe -cuda -allow-multiple
sd.gui.exe -avx512 -allow-multiple
    The first instance no need to start with the "-allow-multiple" parameter to run another new one.
• currently the logging is set to debug in the log on the bottom. Are logs are now written in a log file into the appdata directory
  New sd.cpp branch appeared, but when i tested i found some bugs which is prevents to upgrade.

preview6 - 0.1.4

These preview release have some gui enhancements and inner thread signaling / data handling refactoring. This may cause unexpected behaviours. If you experience it, please report it in a new bug on the github's issues page.

Added features

• CivitAi model info and model images previews for embeddings too
  

• Thumbnails are cached wich downloaded from CivitAi info or generated by diffusion

• Queue jobs more details and features per image:
  

• Queue item right click features:
  

• Incompatible controls are disabled respect to the generation mode, eg. on upscaler tab the CFG is ignored

• Added splash screen while application startup and loading

• Only one instance allowed to run in one time
  

• Pending queue items now are possible to pause. If you set to paused, then it just ignored from the queue. Running job can not be paused (not implemented in sd.cpp)

• Paused queue items now possible to set to pending
  

• On a queue item a Requeue clones the job, but set the Status to paused (so not started automatically)
  

• Added semi-auto versioning (see the cube button next to the Settings and Refresh button)

• Added some place holder images to the ui

Bug fixes

• after a txt2img job a new img2img job causing segfault
• on model tab the model filters are works with text filter too
• refresh models now works with the model list (on the Model tab)
• some json converter got more validation (as always)
• more data check on events (eg. data table index check)

Update from previous preview versions

Just run the attached installer. Stop all instances before running the installer.
Queue job files (.json) partially compatible with the new format. May the old jobs images visible as "deleted".
If the application won't start after upgrade, please remove all .json files

Current limitations / known bugs

• large textual embeddings are causing crash

preview5 - 0.1.2beta

These preview release have some gui design refactoring and inner thread signaling refactoring. This may cause unexpected behaviours. If you experience it, please report in a new bug on the github's issues page.

Added features

• CivitAi model info and model image previews for Lora and Checkpoints in the model list
• CivitAi model info and model image previews now will downloading after the model's hash check finished using own thread, so the window not freezee while it's downloading
• App icon on Windows
• UI redesign, resizable window
• On job queue list now the generated and input images are visible, with double click it's opens in the OS default app
• 10 char length sha256 (Aka Autov2 hash)
• updated readme
• updated the backend to the latest stabble-diffusion.cpp

Bug fixes

• after an upscale job a new job causing segfault
• on model tab the search is now faster, but some bug still in at filter
• better "Refresh models" button behavior - but in model list it's make some weird things

Update from previous preview versions

Just run the attached installer. Stop all instances before running the installer. Currently the config only compatible with the preview3 and preview4 versions. Older versions config files are incompatible, need to delete before run the app. If you experience freezee on startup, please delete all .json file witch is generated by the app.

Current limitations / known bugs

• large textual embeddings are causing crash

preview4

Added features

• ESRGAN upscaler (more details here)

Upscaler limitations

• only 4x upscaling available (sd.cpp limitation), but you can upscale an upscaled image (eg. drag and drop into the "Upscale" tab the already upscaled image :) )
• only tested with RealESRGAN_x4plus_anime_6B.pth. The RealESRGAN_x4plus may work too (with a lot of warning log)
• the upscaler working as tiled upscaler, so tile artifacts sometimes visible

Update from previous preview versions

Just run the attached installer. Stop all instances before running the installer. Currently the config only compatible with the preview3 version. Older versions config files are incompatible, need to delete before run the app.

Current limitations / known bugs

• with controlnet, the base model have to be reloaded on all new image generation (it's automatic at least)

• max 75 tokens in all prompts, more tokens are just ignored

• when lora is specified in the prompt from the previous job, and base model not changed, the old lora keep in context

• when textual embedding is specified in the prompt from the previous job, and base model not changed, the previous textual embedding keep in context

• CivitAi info only works on checkpoints

• CivitAi info downloading is working on the foreground, the window freezee

• SD.next using (10char length sha256) -- ??

preview3

Added features

Job management (Jobs and Images tab)

• double click will show some details about the job (in an ugly window. You can resize it if you want :) )
• now is possible to delete a job (depends on his status) Only one job at once. Just hit right click on it and select "Delete"
• With right click on a queue item you can:
  • Requeue: this will make a clone and start the job
  • Copy to text2img: copy some parameters from the job into the current settings (no prompts will be copied)
  • Copy prompts to ... : copy the prompts from the queue job
  • Detail...: open up the same ugly window as double click. Here you can find some details about the job (nothing fancy, nothing functionality at)
• Job items now sortable by date

GUI enhancements

• re-added desktop notifications, configurable from the settings (turn on/off, set timeout)
• addedd drag and drop functions:
  • text2img accept image drop-in (drag&drop). It can detect meta info from jpeg (which jpeg was generated by this gui or the SD.next)
  • img2img accept image drop-in (drag&drop). It can detect meta info from jpeg (which jpeg was generated by this gui or the SD.next)
• The new default output image format is now jpeg with quality 95. The output images will now store the meta-info from the parameters of the image generation. This images can be drop-in into eg. img2img. It will autofill the prompts and other settings from the image. Some enhancments requied here too. I didn't tested, but in theory, this is compatible with SD.next.

Model management (Models tab)

• Added two checkbox: Lora & Checkpoints
  • This will filter when you use the searchbox
• Added searchbox which will filter the list by name (min. 3 char required to work). If the Lora checkbox is checked but Checkpoints not, then only Lora will be filtered. Need more enhancements :)
• You can sort the models by name
• Loras now in the model list too
• right click:
  • Calculate hash: calclute the sha256 hash from the file. The hash is needed to identify the model by CivitAi. (currently the hash is not compatible with Automatic and other webuis)
  • Select model ... : this will select the model on the top list. When hit the "Queue" button somewhere, then this model will be used
• Doubleclick:
  • On checkpoint type model, a new window will popup with some model info from CivitAi. This currently only works with Checkpoints. With lora, it wil freezee the gui and drop a segfault.
  • Currently the model info is not optimized, so the window will freezee while downloading the infos and the images. Please use with cauction

Current limitations / known bugs

• with controlnet, the base model have to be reloaded on all new image generation (it's automatic at least)
• max 75 tokens in all prompts, more tokens are just ignored
• when lora is specified in the prompt from the previous job, and base model not changed, the old lora keep in context
• when textual embedding is specified in the prompt from the previous job, and base model not changed, the previous textual embedding keep in context
• On start, the gui has no default selected quantation (this is not interfere with the normal working)
• CivitAi info only works on checkpoints
• CivitAi info downloading is working on the foreground, the window freezee
• on job start if the selected model need hash, the hashing is starting, but progressbar not upgraded in the Model tab.
• SD.next using (10char length sha256) -- ??

If you have an earlier preview release and the gui won't start, please delete all settings / presets / model info files (all is in .json files) and try again.

preview2

Current features:

• text to image
• image to image
• keep model in the memory between generations (see limitations)
• controlnet
• optional tiled vae
• optional TAESD encoder
• optional VAE model
• custom quantation, model converting on the fly
• save / load / delete presets
• job queue
• progressbar
• desktop notifications
• lora support in the prompts (see limitations) lora:lora_filename_without_extension:weight
• textual embedding support in prompt (see limitations)
• auto generate sha256 from model files
• manual sha256 generation from model files and lora files in model tab
• clone a job
• copy prompt from a job
• copy settings from a job

Current limitations / known bugs

• with controlnet, the base model have to be reloaded on all new image generation (it's automatic at least)
• max 75 tokens in all prompts, more tokens are just ignored
• when lora is specified in the prompt from the previous job, and base model not changed, the old lora keep in context
• when textual embedding is specified in the prompt from the previous job, and base model not changed, the previous textual embedding keep in context
• On start, the gui has no default selected quantation (this is not interfere with the normal working)

preview1

Current features:

• text to image
• image to image
• keep model in the memory between generations (see limitations)
• controlnet
• optional tiled vae
• optional TAESD encoder
• optional VAE model
• custom quantation, model converting on the fly
• save / load / delete presets
• job queue
• progressbar
• desktop notifications
• lora support in the prompts (see limitations) lora:lora_filename_without_extension:weight
• textual embedding support in prompt (see limitations)

Current limitations / known bugs

• with controlnet, the base model have to be reloaded on all new image generation (it's automatic at least)
• max 75 tokens in all prompts, more tokens are just ignored
• when lora is specified in the prompt from the previous job, and base model not changed, the old lora keep in context
• when textual embedding is specified in the prompt from the previous job, and base model not changed, the previous textual embedding keep in context
• if base model changed and lora used in previous job's prompt, the gui is just crashing