changelog.txt

SD GUI 1.12.0 Changelog:
- New: Support for SDXL via Diffusers (Minimum VRAM untested; might need 16 GB for now)
- New: Support for SDXL refiner model, with variable refine strength (or use 0 to disable and only use base model)
- New: Added samplers: DPM++ 2M SDE, DPM++ 2M SDE Karras, UniPC
- New: Custom samplers now work with all implementations including SDXL, ONNX, InstruxtPix2Pix
- New: Added download buttons for SD 1.5, SD 1.5 ONNX, and SDXL Base/Refiner in Dev Tools
- New: Model Converter now supports SDXL Safetensors -> Diffusers conversion
- New: Interval syntax is supported for LoRA weights, e.g. `0.2 > 1.0 : 0.1` will run 0.2-1.0 in 0.1 steps
- Improved: You can now type a resolution manually into the dropdowns, it will automatically get rounded to a valid number
- Improved Model Converter compatibility for certain model types
- Improved: Image Viewer now also shows LoRA weight of generated image, not just name
- Improved: Inpainting mask is no longer auto-cleared when a new image is loaded, now only if its size is different
- Improved: All Diffusers-based implementations can now be canceled without killing the whole process
- Fixed: HuggingFace model downloader was using an incorrect path
- Fixed unhandled error when using an invalid LoRA training project name
- Fixed various bugs


SD GUI 1.11.0 Changelog:
- New: LoRA and LyCORIS support
- New: LoRA training GUI (replaces Dreambooth, works with 8 GB GPUs)
- New: Option to auto-delete generated images afterwards (if they weren't copied to favs)
- Improved: Updater is much more stable now, and only re-installs updated/new packages
- Improved: External VAEs now also work with Diffusers models
- Improved: Image progress bar now takes high-res fix into account instead of running twice
- Improved: Max seed increased from 4294967295 to 9223372036854775807 (int64 max)
- Improved: VAE folders are now separate from model folders (again)
- Improved: Embeddings folder path can now be changed
- Improved: Progress bar and ETA now accounts for queued tasks instead of running per-task
- Improved: Image Viewer buttons (e.g. prev/next) are now disabled when there are no images to show
- Improved: Various UI/UX improvements
- Fixed issues caused by moving the installation folder, this should now work without problems
- Fixed: Applying CLIP Skip to a model would make it impossible to disable it again
- Fixed: Inpainting model name check is no longer case-sensitive and no longer requires a hyphen
- Fixed: "Append To Prompt" button for embeddings will no longer add a comma if there already is one


SD GUI 1.10.0 Changelog:
- New: Support for Stable Diffusion 2.x models
- New: Added outpainting (expanding images into any direction)
- New: Built-in Self-Updater (experimental)
- New: Added Karras samplers, including DPM++ 2M Karras
- New: Image generation symmetry options
- New: Utility to directly download Huggingface Diffusers models (no need for git clone etc.)
- New: Option to skip last CLIP layers ("CLIP Skip") which can improve quality with some models
- New: Chroma key aka "Greenscreen" option when loading an image (e.g. turn black pixels transparent)
- New additional config file "settings.ini" where power users can adjust certain limits (e.g. max res)
- Improved: Loading transparent images from clipboard is now possible
- Improved: Any model format can now be loaded directly, Diffusers (fastest), Safetensors, Ckpt
- Improved: Prompts can now be longer than 77 tokens (experimental)
- Improved: Textual Inversion embeddings are now properly integrated, newer formats supported
- Improved: InstructPix2Pix and SD ONNX no longer need to reload the model on each run
- Improved: Maximum resolution and prompt history limit have been increased
- Improved: UI is now split into categories that can be collapsed
- Improved: A wide range of UI/UX performance improvements, reduced flickering, etc
- Too many bug fixes to list


SD GUI 1.9.1 Changelog:
- New: Resolution can now also be adjusted in InstructPix2Pix mode (resizes input image)
- New: Resulting aspect ratio is shown next to resolution option
- New: Right-click image menu in InstructPix2Pix has an option to copy a comparison image
- Improved: Model compatibility, any Safetensors file should now work fine after conversion
- Improved: Model converter no longer includes 1 GB safety checker in ONNX outputs
- Improved: Image Viewer now shows image resolution info in InstructPix2Pix mode
- Improved: InstructPix2Pix no longer downloads unnecessary files, now <2GB
- Fixed: InstructPix2Pix errors when trying to run it without internet connection
- Fixed: Certain options result in filenames being too long causing issues
- Fixed: Stable Diffusion model gets reloaded if amount of free RAM changed
- Fixed: Image export breaks if initialization image(s) import takes too long
- Fixed: "Open Output Folder" opens Documents folder
- Fixed: ONNX and InstructPix2Pix don't accept images with UTF8 characters in path
- Fixed: Running InvokeAI post-processing sometimes fails due to long paths
- Fixed: Issues with comma vs dot, depending on Windows region settings


SD GUI 1.9.0 Changelog:
- New: Added InstructPix2Pix (Use with Settings -> Implementation -> InstructPix2Pix)
- New: Added the option to show the input image next to the output for comparisons
- New: Added option to choose output filename timestamp (None, Date, Date+Time, Epoch)
- Improved: minor UI fixes, e.g. no more scrollbar in main view if there is enough space
- Fixed: Dreambooth learning rate multiplier would default to 2x
- Fixed: Minor PNG metadata parsing issues
- Fixed: Various of other minor fixes


SD GUI 1.8.1 Changelog:
- New: RunwayML inpainting now also works with ONNX implementation
- New: Version now checks for the latest version on startup (no updater yet)
- Improved: High-Res Fix can now be enabled if any axis is >512px instead of both
- Improved: High-Res Fix option is saved/loaded when closing and re-opening GUI
- Improved: Inpainting Mask Blur is now automatically disabled when using RunwayML inpainting
- Improved: In Installer, a custom git commit can now be used (for developers)
- Fixed: Upscalers were disabled by default on <=6GB GPUs
- Fixed: Model loading would fail without an internet connection
- Fixed: ONNX seeding did not work
- Fixed: CFG Scale <=1 didn't work or would fallback to default value
- Fixed: Inpainting mask was saved with irreversible blur, making editing harder
- Fixed: Init image import would ignore stretch/pad setting
- Fixed: Saved inpainting mask wouldn't get cleared properly when resetting it
- Fixed: Various minor bugs


SD GUI 1.8.0 Changelog:
- New: Stable Diffusion DirectML implementation, enables image generation on AMD GPUs
- New: Inpainting models are now supported, providing much better quality than the old method
- New: Models are cached in RAM when switching. Switching back loads a model in ~2 seconds
- New: Text-based masking - Describe what you want to mask instead of drawing the mask
- New: Initialization image can be previewed by Shift+Hovering over the filename in the GUI
- New: Generation resolution is automatically set to init image size if it doesn't match
- New: Pop-Up Image Viewer now has an option to enable Slideshow Mode by default
- New: Model converter (Dev Tools) to convert between Pytorch, Diffusers, ONNX, Safetensors
- New: Image Import window has a new option to use metadata plus the image itself as init image
- Improved: Image Viewer now has hotkeys Ctrl+Left/Ctrl+Right for previous/next image
- Improved: Image Viewer hotkeys (copy, open, etc.) now also work in pop-up viewer
- Improved: Mask Editor can be resized, is no longer locked to the init image size
- Improved: Mask Editor now has paste, invert, save, load, and undo functionality
- Improved: Inpainting mask can now be edited without having to clear it first
- Improved: Prompt History will not save a new entry if it's identical to the previous one
- Improved: Prompt History now has a limit of 500 entries to avoid long loading times
- Improved: Prompt Queue now behaves like a queue, not like a stack (FIFO)
- Improved: Welcome Screen (MOTD) can now be disabled after it's been shown once per version
- Improved: Image Import window now shows resolution, import actions are a dropdown to avoid clutter
- Improved: Mask Editor performance is better now, blur is only applied after releasing mouse
- Improved: User can now ignore the Dreambooth VRAM warning
- Fixed: Post-processing after generation would cause a soft-lock
- Fixed: Loading settings from prompt history sets incorrect values
- Fixed: Broken or malicious model files would cause an unhandled exception
- Fixed: Using a concept with <Name> syntax would show warning about missing placeholder
- Fixed: InvokeAI CLI wouldn't open if GUI image generation was cancelled before
- Fixed: Unhandled error when trying to use init images that no longer exist
- Fixed: GUI would freeze if no model files exist in models folder(s)


SD GUI 1.7.2 Changelog:
- Improved prompt syntax converter (converts A1111 or pre-1.7 weighting to new syntax)
- VAE Model selection now also lists .pt files (same format as ckpt, just different extension)
- Fixed init images being in random order after importing
- Fixed idle CPU usage
- Fixed an issue with having the same model filename in multiple model folders
- Fixed issue where k_diffusion module was not imported properly
- If a config YAML exists next to a model file it will be used instead of the default (e.g. model.ckpt.yaml)


SD GUI 1.7.1 Changelog:
- Added DPM++2 and DPM++2A samplers
- Added advanced syntax for step count (run same image with different step counts)
- New prompt attention syntax - Use (text)++ instead of ((text)), etc.
- Added automatic conversion of old attention syntax to new syntax
- Security: Added pickle malware scanner (checks model files automatically)
- Minor UI improvements, DreamBooth now has a dedicated button
- Fixed issues with pytorch not working correctly (No CUDA GPU detected)
- Fixed issues with post-processing after generation
- Fixed negative prompt input being visible when using Low Memory Mode
- Fixed certain sliders loading invalid values after restarting the program
- Reduced installation size and optimized code


SD GUI 1.7.0 Changelog:
- Added support for running multiple init images at once, e.g. for animation frames
- Added support for wildcards (insert words/phrases into prompt dynamically)
- Added "Hi-res Fix" which allows better results at higher (>512px) resolutions
- Added support for loading custom VAE models (can improve image quality)
- Added support for running upscaling or face restoration manually
- Added support for all samplers when using an init image (not just DDIM)
- Added button to resize the window to fit the currently displayed image
- Added option to save output images in a subfolder per session
- Added options to use seamless mode only on one axis (horizontal or vertical)
- Added a label that displays the current image's prompt
- Added hotkeys: Quick-switch VAE, copy current image, copy image to favorites
- Added new developer tool: Open CMD in SD Conda environment
- DreamBooth training: Images can now be automatically resized if they are not 512x512
- DreamBooth training: Added slider for steps multiplier (0.5x-2x)
- GUI improvements, prompt and negative prompt are now separate
- Image viewer now also shows "actual" image resolution (for upscaled images)
- Sliders now also allow you to type a value instead of dragging the handle
- Loading image metadata now also works for images generated with automatic1111
- "Delete All Current Images" now requires a confirmation
- Improved prompt history/queue UI (show full prompt on hover, and more)
- Improved GUI rendering, should have less flickering now
- Fixed bug where model pruning would say "Failed..." even if it was successful


SD GUI 1.6.0 Changelog:
- Added Dreambooth training (24 GB VRAM required for now!)
- Added support for prompt emphasis - () for more, {} for less
- Added Model Quick Switcher: Press Ctrl+M to quickly change the current AI model
- Added model folder manager: You can now add additional model folders to load models from
- Pop-Up Image Viewer: When slideshow mode is enabled, Left/Right arrow keys change images
- Pop-Up Image Viewer: Window can now be resized/zoomed in 25% steps using mouse wheel
- Pop-Up Image Viewer: Added "Always on Top" option to keep window always on top
- Added lots of hotkeys, documented on Github Readme
- Words in prompt field can now be deleted using CTRL+BACKSPACE, like in most text editors
- Model Pruning Tool: Added option to delete input file if pruning was successful
- Fixed an issue where the Stable Diffusion process would be killed when cancelling
- Fixed prompt queue not working after running first entry


SD GUI 1.5.0 Changelog:
- Upstream Code Update: Supports exclusion words, runs on 4 GB VRAM (when no other apps are open)
- UI is now more flexible, window can be resized, prompt field is bigger and has zoomable text
- Added CodeFormer face restoration as an alternative to GFPGAN
- Updated RealESRGAN (upscaler), should now be faster with same or better quality
- Added button to delete either the current image, or all generated images
- Added separate checkboxes to choose if you want to include prompt/seed/scale/sampler/model in filename
- Added option to save original image in addition to the post-processed image (if post-proc is enabled)
- Added option to select the CUDA device (Automatic, CPU, or a specific GPU)
- Added model merging tool
- Added model pruning tool (strip EMA data and/or convert to fp16 half-precision for 2 GB models)
- Added option to unload Stable Diffusion after each generation (like in pre-1.4.0)
- Added reliable orphan process handling (Python no longer stays in RAM if the GUI crashes)
- Image Viewer: Added short cooldown after using prev/next image buttons, before the newest will be shown again
- Image Viewer context menu: Added button to re-generate single image with current settings/seed
- Image Viewer Pop-up: Now borderless, 100% zoom by default, double-click for fullscreen
- Image Viewer Pop-up: Added "Slideshow Mode" which mirrors the regular image viewer when enabled
- Added image load form, allows you to use as init image, load settings from metadata, or copy prompt
- Images can now be loaded from clipboard, not just from files
- Prompt History: Added option to disable history, added text filter
- You can now add an entry to the prompt queue by right-clicking on its icon
- Disabled post-processing with Low Memory Mode as it was not working properly
- Prompt text in folder/file names now strips weighting (won't create new folders for each weight change)
- Current model name gets printed whenever Stable Diffusion is started
- Full Precision is now enabled by default on GTX 16 series cards to fix compatibility with them
- Fixed empty/invalid prompts (e.g. newlines) counting towards the target image amount
- Some fixes regarding cancelling the generation process and handling crashes

NOTES:
Low Memory Mode is a low priority for me because it's a separate codebase so adding features is hard.
Also, the regular mode can now run on 6 GB easily, and even 4 GB if all other GPU apps are closed.
Apart from that, it's now possible to run the regular mode on CPU, which is slow, but it works.


SD GUI 1.4.0 Changelog:
- Stable Diffusion model no longer needs to be reloaded every time new images are generated
- Added support for mask-based inpainting
- Added support for loading HuggingFace .bin concepts (textual inversion embeddings)
- Added prompt queue, allows you to queue up prompts with their settings
- Added prompt history, allows your to view or load previous prompts
- Added a progress bar that shows the generation progress of the current image
- Added option to play sound and/or show notification when finished
- Added a "Welcome screen" that shows the changelog and patrons
- Added buttons to use the previous seed or to reset it
- Added button to expand prompt field to 2x height
- Added proper support for DPI scaling
- Post-processing now works with Low Memory Mode
- Further VRAM optimizations, especially in regular mode
- Available CUDA GPUs are now listed in window title
- Windows should not be able to go to sleep while the program runs (untested)
- Updated GFPGAN model to 1.4
- Fixed a bug where empty "unknown prompt" folders were created
- Fixed some issues regarding the python environment


SD GUI 1.3.1 Changelog:
- Init image (img2img) now works with Low Memory Mode
- Log button now has options to copy each log directly
- Added Stable Diffusion model selection in Settings
- Added a right-click menu to the full-screen image viewer window with "Copy Image" option
- Added warning if prompt might be too long for CLIP (>70 words)
- Installation should no longer conflict with system Python/Conda installations
- Installer will no longer re-install Python environment by default (if it already exists)
- Fixed output folder not being created if it does not exist
- Fixed file type check (for init image, embeddings) being case-sensitive
- Fixed uppercase file extensions not being accepted for init images or embeddings
- Fixed original image always being saved in addition to upscaled image
- VRAM usage optimizations
- Other minor bug fixes


SD GUI 1.3.0 Changelog:
- Added option to use low-VRAM code, only needs around 3.3 GB VRAM for 512x512 images
- Added optional AI post-processing: RealESRGAN upscaling and GFPGAN face restoration
- Added option to generate seamless (tileable) images
- Added fullscreen image preview (click on image to open), click into it for 2x2/3x3 tiling (if seamless mode is enabled)
- Added option to save images to a subfolder per prompt
- Added option to disable prompt in filename
- User can now run multiple prompts at once (one prompt per line, not counting word wrap)
- Added option to use the same seed for each prompt when running multiple prompts
- Added option to change image output folder
- Added warnings if the program is running from a long or problematic (e.g. OneDrive) path
- Slightly reduced VRAM usage across both implementations (fast/optimized)
- Improved installer, python files are now included and should no longer conflict with system conda
- Error messages are now shown for common errors
- Fixed full-precision option not actually enabling full-precision
- Fixed DPI scaling breaking
- General logging and UX improvements
- UI improvements