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