Release v0.6.0

• SDXL support
• Removed original bislerp
• FlashAttention
• Better clip skip
• Automatic model version detection
• SAG rework
• Updated default settings for lower VRAM usage
• Settings difference resolver (migration to newer versions should be smoother)
• Partial support for float8 (only for storing models in VRAM, but drastically reduces VRAM usage)
• Heun++ sampler
• Kohya deepshrink
• Scalecrafter
• Added CODEOWNERS
• Updated CivitAI browser UI
• SASolver
• General postprocessors (hi-res, upscales, etc.)
• UI improvements for mobile users (WIP)
• FreeU v2 support, default FreeU v2 settings picker
• ADetailer
• More predictable model loading

• Highres fix can now use image upscalers (ESRGAN, RealSR, etc.) for the intermediate step
• API converted to sync where applicable, this should resolve some issues with websockets and thread lockups
• FreeU support
• Diffusers version bump

Critical bug fix

• Automatic installer now correctly installs CUDA versions of PyTorch

Changelog

v0.4.0

Biggest Changes

• Hi-res fix for AITemplate
• Model and VAE autoloading
• Partial support for Kdiffusion samplers (might be broken in some cases - controlnet, hi-res...)
• Hypertile support (https://github.com/tfernd/HyperTile)
• Theme overhaul

All Changes

• Added docs for Vast.ai
• Better Settings UI
• Updated Model Manager
• Garbage Collection improvements
• Hi-res fix for AITemplate
• Added DPMSolverSDE diffusers sampler
• Model autoloading
• Partial support for Kdiffusion samplers (might be broken in some cases - controlnet, hi-res...)
• Tag autofill
• New SendTo UI that is context-aware
• Fixed symlink deletion bug for models
• New documentation theme
• New sigma types for Kdiffusion (exponential, polyexponential, VP)
• Image upload should now display the correct dimensions
• Fixed WebP crashes in some cases
• Remote LoRA support <lora:URL:weight>
• Fix some CORS issues
• Hypertile support (https://github.com/tfernd/HyperTile)
• Fixed uvicorn logging issues
• Fixed update checker
• Added some extra tooltips to the UI
• Sampler config override for people who hate their free time
• Bumped dependency versions
• Image browser entries should get sorted on the server, removing the need for layout shift in the UI
• Cleaned up some old documentation
• Transfer project from PyLint to Ruff
• GitHub Actions CI for Ruff linting
• Theme overhaul
• Fixed NaiveUI ThemeEditor
• Sort models in the Model Loader
• Console logs are now accessible in the UI
• ...and probably a lot more that I already forgot

Contributors

• gabe56f (https://github.com/gabe56f)
• Stax124 (https://github.com/Stax124)
• Katehuuh (https://github.com/Katehuuh)

Release v0.3.1

• Patched Hires fix
• Patched lycoris path bug

Release v0.3.0

Most notable changes

• Dynamic AITemplate (with fixed ControlNet)
• Reworked LoRAs and support or LyCORIS
• CivitAI browser
• Arbitrary image size (non multiples of 8 should now work fine)
• Support for local upscalers
• Settings autosave
• AIT support for long weighted prompts
• CLIP skip
• VAE switching
• Frontend SPA routing fixed
• API support for sending images as Bytes
• Support for WebP and JPEG file-formats
• HF Token no longer needed
• Fractional SD upscaling
• Karras Sigmas for KDPM schedulers
• ControlNet 1.1 support

...and a lot of bugfixes, UI changes, purging of old code.

Diff: https://github.com/VoltaML/voltaML-fast-stable-diffusion/compare/v0.2...v0.3.0

Big thanks to all contributors ❤️

Release v0.2

Attention

I misspelled "SDPA" as "SPDA", so, if you used this attention, you would need to set it to the fixed one in the UI and click Save Config to use it again.

Changelog

• ControlNet now has the option to run with a preprocessed image
• Self -attention guidance -> may produce better images, but reduces speed
• LoRA weights settings
• Better defaults
• Model upload from the local system to host
• Flashbang mode (Light mode) - I am sorry, I truly am
• Model manager - only allows model deletion, but will be extended in the future
• 9-channel inpainting
• New format for logger
• AMD GPU support for PyTorch
• DirectML Support
• Spinners for a better indication that your PC hasn't frozen, but is on the edge of exploding
• Fixed ControlNet dimensions
• Tiled RealESRGAN
• Automatic Textual Inversion loader
• Fixed Deepdanbooru threshold
• Keyboard shortcuts (CTRL+Arrow Up/Down to change attention inside a textbox, CTRL+Enter to generate an image)
• Better Queue system
• Continuous image generation mode for faster GPUs (generate an image on change in prompt/params)
• Enabled Scribble ControlNet
• Tagger and Upscaler now work without other models loaded
• Initial support for Dynamic AIT in place (but still under heavy development and only accessible through API)
• Fixed Upscaler not saving images properly
• Allowed user to change a port of the app
• Transitioned Docs from Algolia to locally generated search for more up-to-date results
• Default attention set to SDPA by default
• NaiveUI Theme editor support - feel free to customize how Volta looks (needs to be enabled in the settings first)
• Bumped dependency versions
• Affixed dependency versions for a better experience
• Support for Subquadratic and Multihead attentions
• Added Bislerp interpolation option for Hi-res
• Output folder and format can now be changed in the settings
• New Image Browser
• Fixed some spelling in docs
• Fixed dotenv not being applied properly
• AITemplate cache is now automatically wiped for every compile

Roadmap

• Finish the documentation
• Dynamic AIT
• More UI improvements and cleanups

Changelog

• Support for TomeSD
• AIT ControlNet caching
• UI shows the last seed used
• "Send to" UI feature
• "Image Browser" Delete Image, Download image buttons
• The "interrupt" button will be disabled if an image is not being generated
• Image to Image steps should now have the correct number of total steps
• Deepdanbooru tagger
• Txt2img High-resolution fix
• New Image Upload component (D&D support)
• AIT VRAM savings (should be much closer to normal diffusers pipeline)
• AIT support for .safetensors/.ckpt
• Can now run on CPU (and maybe MPS - not tested as we lack the hardware)
• LoRA support for PyTorch
• Much more robust UI
• ControlNet AIT
• full WSL support (recommended over Docker)
• Full support for .safetensors/.ckpt
• Much much more

I am also in the process of developing a cross-platform manager called "VoltaML Manager". It is a Rust-based utility that should take care of the installation process, updating, and gathering of diagnostic data so you do not have to do it on your own (not telemetry - just a local process, that you can then take the output and paste it into the Issue context)

The source code for that feature can be found here: https://github.com/VoltaML/voltaML-fast-stable-diffusion/tree/experimental/manager

Next up

• Finally add more upscalers and redo the UI for them (RealESRGANs, Valar4x, and so on)
• 9-channel inpainting with a better inpainting UI
• ControlNet 1.1, MultiControlNet
• Control over LoRA bias weights
• Text Embeddings
• Presets in the UI (1920x1080, 512x836...)
• Redo the reconnect Button with more functionality
• Light theme UI (accessibility reasons)

Release 0.0.4 - Merger with experimental

Main changes

• New Frontend written in Vue3 (+NaiveUI)
• Switched from Flask to FastAPI
• PyTorch 2.0
• SPD Attention (https://huggingface.co/docs/diffusers/v0.13.0/en/optimization/torch2.0#fp16-benchmark)
• Project Documentation (https://voltaml.github.io/voltaML-fast-stable-diffusion/)
• Automated CI/CD pipeline (will get better with time)
• TensorRT was deprecated in favor of AITemplate (that allows us to use new diffusers and transformers)
• DevContainer one click no dependencies development
• Discord Bot
• Support for local installation (PyTorch only)
• Much cleaner code-base
• Support for multiple models
• Support for multiple GPUs
• Documented and Strong Typed API (thanks to FastAPI)