- Added Reference-Only mode for Inference ControlNet, used for guiding the sampler with an image without a pretrained model. Part of the latent and attention layers will be connected to the reference image, similar to Image to Image or Inpainting.
- Inference Primary Sampler Addons (i.e. ControlNet, FreeU) are now inherited by Hires Fix Samplers, this can be overriden from the Hires Fix module's settings menu by disabling the "Inherit Primary Sampler Addons" option.
- Inference ControlNet module now supports over 42 preprocessors, a new button next to the preprocessors dropdown allows previewing the output of the selected preprocessor on the image.
- Added resolution selection for Inference ControlNet module, this controls preprocessor resolution too.
- Added new package: [StableSwarmUI](https://github.com/Stability-AI/StableSwarmUI) by Stability AI
- Added new package: [Stable Diffusion WebUI Forge](https://github.com/lllyasviel/stable-diffusion-webui-forge) by lllyasviel
- Added extension management for SD.Next and Stable Diffusion WebUI-UX
- Added the ability to choose where CivitAI model downloads are saved
- Added `--launch-package` argument to launch a specific package on startup, using display name or package ID (i.e. `--launch-package "Stable Diffusion WebUI Forge"` or `--launch-package c0b3ecc5-9664-4be9-952d-a10b3dcaee14`)
- Added more Base Model search options to the CivitAI Model Browser
- Added Stable Cascade to the HuggingFace Model Browser
#### Inference
- Added Inference Prompt Styles, with Prompt Expansion model support (i.e. Fooocus V2)
- Added option to load a .yaml config file next to the model with the same name. Can be used with VPred and other models that require a config file.
- Added copy image support on linux and macOS for Inference outputs viewer menu
- Added `--launch-package` argument to launch a specific package on startup, using display name or package ID (i.e. `--launch-package "Stable Diffusion WebUI Forge"` or `--launch-package c0b3ecc5-9664-4be9-952d-a10b3dcaee14`)
- Fixed repeated nested folders being created in `Models/ControlNet` when using ComfyUI in Symlink mode. Existing folders will be repaired to their original structure on launch.
- Changed Symlink shared folder link targets for Automatic1111 and ComfyUI. From `ControlNet -> models/controlnet` to `ControlNet -> models/controlnet/ControlNet` and `T2IAdapter -> models/controlnet/T2IAdapter`.
- Use the plus button to add new steps (Hires Fix, Upscaler, and Save Image are currently available), and the edit button to enable removing or dragging steps to reorder them. This enables multi-pass Hires Fix, mixing different upscalers, and saving intermediate images at any point in the pipeline.
- Added Sampler addons
- Addons usually affect guidance like ControlNet, T2I, FreeU, and other addons to come. They apply to the individual sampler, so you can mix and match different ControlNets for Base and Hires Fix, or use the current output from a previous sampler as ControlNet guidance image for HighRes passes.
- Added SD Turbo Scheduler
- Added display names for new samplers ("Heun++ 2", "DDPM", "LCM")
- Lykos Account sign-up and login - currently for Patreon OAuth connections but GitHub requests caching and settings sync are planned
- Supporters can now connect your Patreon accounts, then head to the Updates page to choose to receive auto-updates from the Dev or Preview channels
- CivitAI Account login with API key - enables downloading models from the Browser page that require CivitAI logins, more integrations like liking and commenting are also planned
#### Updates Settings Subpage
- Toggle auto-update notifications and manually check for updates
- Choose between Stable, Preview, and Dev update channels
#### Inference Settings Subpage
- Moved Inference settings to subpage
- Updated with more localized labels
#### Outputs Page
- Added Refresh button to update gallery from file system changes
#### Checkpoints Page
- Added the ability to drag & drop checkpoints between different folders
- Added "Copy Trigger Words" option to the three-dots menu on the Checkpoints page (when data is available)
- Added trigger words on checkpoint card and tooltip
- Added "Find Connected Metadata" options for root-level and file-level scans
- Added "Update Existing Metadata" button
#### Model Browser
- Added Hugging Face tab to the Model Browser
- Added additional base model filter options for CivitAI ("SD 1.5 LCM", "SDXL 1.0 LCM", "SDXL Turbo", "Other")
- Added the ability to type in a specific page number in the CivitAI Model Browser
- Right clicking anywhere on the model card will open the same menu as the three-dots button
- New model downloads will save trigger words in metadata, if available
- Model author username and avatar display, with clickable link to their profile
### Changed
#### General
- Model Browser page has been redesigned, featuring more information like rating and download counts
- Model Browser navigation has improved animations on hover and compact number formatting
- Updated Outputs Page button and menu layout
- Rearranged Add Package dialog slightly to accommodate longer package list
- Folder-level "Find Connected Metadata" now scans the selected folder and its subfolders
- Model Browser now split into "CivitAI" and "Hugging Face" tabs
#### Inference
- Selected images (i.e. Image2Image, Upscale, ControlNet) will now save their source paths saved and restored on load. If the image is moved or deleted, the selection will show as missing and can be reselected
- Project files (.smproj) have been updated to v3, existing projects will be upgraded on load and will no longer be compatible with older versions of Stability Matrix
### Fixed
- Fixed Outputs page reverting back to Shared Output Folder every time the page is reloaded
- Potentially fixed updates sometimes clearing settings or launching in the wrong directory
- Improved startup time and window load time after exiting dialogs
- Fixed control character decoding that caused some progress bars to show as `\u2588`
- Fixed Python `rich` package's progress bars not showing in console
- Optimized ProgressRing animation bindings to reduce CPU usage
- Improved safety checks in custom control rendering to reduce potential graphical artifacts
- Improved console rendering safety with cursor line increment clamping, as potential fix for [#111](https://github.com/LykosAI/StabilityMatrix/issues/111)
- Fixed [#290](https://github.com/LykosAI/StabilityMatrix/issues/290) - Model browser crash due to text trimming certain unicode characters
- Fixed crash when loading an empty settings file
- Improve Settings save and load performance with .NET 8 Source Generating Serialization
- Fixed ApplicationException during database shutdown
- InvokeAI model links for T2I/IpAdapters now point to the correct folders
- Added extra checks to help prevent settings resetting in certain scenarios
- Fixed Refiner model enabled state not saving to Inference project files
- Fixed NullReference error labels when clearing the Inference batch size settings, now shows improved message with minimum and maximum value constraints
- Use the plus button to add new steps (Hires Fix, Upscaler, and Save Image are currently available), and the edit button to enable removing or dragging steps to reorder them. This enables multi-pass Hires Fix, mixing different upscalers, and saving intermediate images at any point in the pipeline.
- Added Sampler addons
- Addons usually affect guidance like ControlNet, T2I, FreeU, and other addons to come. They apply to the individual sampler, so you can mix and match different ControlNets for Base and Hires Fix, or use the current output from a previous sampler as ControlNet guidance image for HighRes passes.
- Selected images (i.e. Image2Image, Upscale, ControlNet) will now save their source paths saved and restored on load. If the image is moved or deleted, the selection will show as missing and can be reselected
- Project files (.smproj) have been updated to v3, existing projects will be upgraded on load and will no longer be compatible with older versions of Stability Matrix
- Fixed control character decoding that caused some progress bars to show as `\u2588`
- Fixed Python `rich` package's progress bars not showing in console
- Optimized ProgressRing animation bindings to reduce CPU usage
- Improved safety checks in custom control rendering to reduce potential graphical artifacts
- Improved console rendering safety with cursor line increment clamping, as potential fix for [#111](https://github.com/LykosAI/StabilityMatrix/issues/111)
- Lykos Account sign-up and login - currently for Patreon OAuth connections but GitHub requests caching and settings sync are planned
- Supporters can now connect your Patreon accounts, then head to the Updates page to choose to receive auto-updates from the Dev or Preview channels
- CivitAI Account login with API key - enables downloading models from the Browser page that require CivitAI logins, more integrations like liking and commenting are also planned
- Updates Settings Subpage
- Toggle auto-update notifications and manually check for updates
- Choose between Stable, Preview, and Dev update channels
- Fixed [#297](https://github.com/LykosAI/StabilityMatrix/issues/297) - Model browser LiteAsyncException occuring when fetching entries with unrecognized values from enum name changes
- This will link the package's output folders to the relevant subfolders in the "Outputs" directory
- When a package only has a generic "outputs" folder, all generated images from that package will be linked to the "Outputs\Text2Img" folder when this option is enabled
- Fixed error `got an unexpected keyword argument 'socket_options'` on fresh installs of Automatic1111 Stable Diffusion WebUI due to missing httpx dependency specification from gradio
- Right click Inference Batch options to enable selecting a "Batch Index". This can be used to reproduce a specific image from a batch generation. The field will be automatically populated in metadata of individual images from a batch generation.
- The index is 1-based, so the first image in a batch is index 1, and the last image is the batch size.
- Currently this generates different individual images for batches using Ancestral samplers, due to an upstream ComfyUI issue with noise masking. Looking into fixing this.
- Added toggle to show connected model images in the Checkpoints tab
- Added "Find Connected Metadata" option to the context menu of Checkpoint Folders in the Checkpoints tab to connect models that don't have any metadata
- Fixed [#97](https://github.com/LykosAI/StabilityMatrix/issues/97) - Codeformer folder should now get linked correctly
- Fixed [#106](https://github.com/LykosAI/StabilityMatrix/issues/106) - ComfyUI should now install correctly on Windows machines with an AMD GPU using DirectML
- Fixed [#59](https://github.com/LykosAI/StabilityMatrix/issues/61) - `GIT` environment variable is now set for the embedded portable git on Windows as A1111 uses it instead of default `PATH` resolution
- Fixed model categories sometimes not showing if they are empty
- Improved model download hash verification performance
- Fixed some text wrapping visuals on expanded model version dialog on model browser
- Added cancel button for create folder dialog
- One click first time installer now defaults to using the "Package" name instead of the display name ("stable-diffusion-webui" instead of "Stable Diffusion WebUI") for the install folder name - probably safer against upstream issues on folder names with spaces.
- Fully rewritten using Avalonia for improved UI and cross-platform support, our biggest update so far, with over 18,000 lines of code.
- Release support for Windows and Linux, with macOS coming soon
- Model Browser now indicates models that are already downloaded / need updates
- Checkpoints Manager now supports filtering/searching
- One-click installer now suggests all 3 WebUI packages for selection
- Hardware compatibility and GPU detection is now more accurate
- Download Indicator on the nav menu for ongoing downloads and progress; supports multiple concurrent model downloads
- Improved console with syntax highlighting, and provisional ANSI rendering for progress bars and advanced graphics
- Input can now be sent to the running package process using the top-right keyboard button on the Launch page. Package input requests for a (y/n) response will now have an interactive popup.
