Stable Diffusion WebUI Forge - Classic

^{[ Classic | Neo ]}

Stable Diffusion WebUI Forge is a platform on top of the original Stable Diffusion WebUI by AUTOMATIC1111, to make development easier, optimize resource management, speed up inference, and study experimental features.
The name "Forge" is inspired by "Minecraft Forge". This project aims to become the Forge of Stable Diffusion WebUI.

- lllyasviel
^{(paraphrased)}

"Classic" mainly serves as an archive for the "previous" version of Forge, which was built on Gradio 3.41.2 before the major changes (see the original announcement) were introduced. Additionally, this fork is focused exclusively on SD1 and SDXL checkpoints, having various optimizations implemented, with the main goal of being the lightest WebUI without any bloatwares.

Tip

How to Install

Features [Dec.]

Most base features of the original Automatic1111 Webui should still function

New Features

• Support uv package manager
  • requires manually installing uv
  • drastically speed up installation
  • see Commandline
• Support SageAttention, FlashAttention, and fast fp16_accumulation
  • see Commandline
• Support fast fp8 operation (torch._scaled_mm)
  • requires RTX 40 +
  • requires UNet Weights in fp8 option
  • ~10% speed up; reduce quality
  • enable in Settings/Optimizations

Note

The fp16_accumulation requires fp16 precision, thus is not compatible with the fast fp8 operation

• Persistent LoRA Patching
  • speed up LoRA loading in subsequent generations
  • see Commandline
• Patch LoRA in-place
  • reduce VRAM usage when loading LoRA
  • enable in Settings/Extra Networks
• Implement new Samplers
  • (ported from reForge Webui)
• Implement Scheduler dropdown
  • (backported from Automatic1111 Webui upstream)
  • enable in Settings/UI Alternatives
• Add CFG slider to the Hires. fix section
• Implement RescaleCFG
  • reduce burnt colors; mainly for v-pred checkpoints
  • enable in Settings/UI Alternatives
• Implement MaHiRo
  • alternative CFG calculation; improve prompt adherence
  • enable in Settings/UI Alternatives
• Implement Epsilon Scaling
  • enable in Settings/Stable Diffusion
• Implement full precision calculation for Mask blur blending
  • enable in Settings/img2img
• Support loading upscalers in half precision
  • speed up; reduce quality
  • enable in Settings/Upscaling
• Support running tile composition on GPU
  • enable in Settings/Upscaling
• Allow newline in LoRA metadata
  • (backported from Automatic1111 Webui upstream)
• Implement sending parameters from generation result rather than from UI
  • e.g. send the prompts instead of Wildcard syntax
  • enable in Settings/Infotext
• Implement tiling optimization for VAE
  • reduce memory usage; reduce speed
  • enable in Settings/VAE
• Implement diskcache for hashes
  • (backported from Automatic1111 Webui upstream)
• Implement skip_early_cond
  • (backported from Automatic1111 Webui upstream)
  • enable in Settings/Optimizations
• Allow inserting the upscaled image to the Gallery instead of overriding the input image
  • (backported from upstream PR)
• Support v-pred SDXL checkpoints (e.g. NoobAI)
• Support new LoRA architectures
• Update spandrel
  • support new Upscaler architectures
• Add support for .avif, .heif, and .jxl image formats
• Automatically determine the optimal row count for X/Y/Z Plot
• Support new LoRA architectures
• DepthAnything v2 Preprocessor
• Support NoobAI Inpaint ControlNet
• Support Union / ProMax ControlNet
  • they simply always show up in the dropdown

Removed Features

• SD2
• Alt-Diffusion
• Instruct-Pix2Pix
• Hypernetworks
• SVD
• Z123
• CLIP Interrogator
• Deepbooru Interrogator
• Textual Inversion Training
• Checkpoint Merging
• LDSR
• Most built-in Extensions
• Some built-in Scripts
• Some Samplers
• Sampler in RadioGroup
• test scripts
• Some Preprocessors (ControlNet)
• Photopea and openpose_editor (ControlNet)
• Unix .sh launch scripts
  • You can still use this WebUI by simply copying a launch script from other working WebUI

Optimizations

• [Freedom] Natively integrate the SD1 and SDXL logics
  • no longer git clone any repository on fresh install
  • no more random hacks and monkey patches
• Fix canvas-zoom-and-pan built-in extension
  • no more infinite-resizing bug when using Send to buttons
• Fix RAM and VRAM leak when switching checkpoints
• Clean up the ldm_patched (i.e. comfy) folder
• Remove unused cmd_args
• Remove unused args_parser
• Remove unused shared_options
• Remove legacy codes
• Fix some typos
• Remove redundant upscaler codes
  • put every upscaler inside the ESRGAN folder
• Optimize upscaler logics
• Optimize certain operations in Spandrel
• Optimize the creation of Extra Networks pages
  • (backported from Automatic1111 Webui upstream)
• Improve color correction
• Improve hash caching
• Improve error logs
  • no longer print TypeError: 'NoneType' object is not iterable
• Improve implementation for Comments
• Update the implementation for uni_pc sampler
• Revamp settings
  • improve formatting
  • update descriptions
• Check for Extension updates in parallel
• Move embeddings folder into models folder
• ControlNet Rewrite
  • change Units to gr.Tab
  • remove multi-inputs, as they are "misleading"
  • change visible toggle to interactive toggle; now the UI will no longer jump around
  • improve Presets application
  • fix Inpaint not masked mode
• Disable Refiner by default
  • enable again in Settings/UI Alternatives
• Disable Tree View by default
  • enable again in Settings/Extra Networks
• Hide Sampler Parameters by default
  • enable again by adding --adv-samplers flag
• Hide some X/Y/Z Plot options by default
  • enable again by adding --adv-xyz flag
• Run text encoder on CPU by default
• Fix pydantic Errors
• Fix Soft Inpainting
• Fix Controllllite
• Fix MultiDiffusion
• Fix SD Upscale
• Lint & Format
• Update Pillow
  • faster image processing
• Update protobuf
  • faster insightface loading
• Update to latest PyTorch
  • torch==2.9.1+cu130
  • xformers==0.0.33

Note

If your GPU does not support the latest PyTorch, manually install older version of PyTorch

• No longer install open-clip twice
• Update some packages to newer versions
• Update recommended Python to 3.11.9
• many more... ™️

Commandline

These flags can be added after the set COMMANDLINE_ARGS= line in the webui-user.bat (separate each flag with space)

A1111 built-in

• --no-download-sd-model: Do not download a default checkpoint
  • can be removed after you download some checkpoints of your choice
• --xformers: Install the xformers package to speed up generation
• --port: Specify a server port to use
  • defaults to 7860
• --api: Enable API access

• Once you have successfully launched the WebUI, you can add the following flags to bypass some validation steps in order to improve the Startup time
  • --skip-prepare-environment
  • --skip-install
  • --skip-python-version-check
  • --skip-torch-cuda-test
  • --skip-version-check

Important

Remove them if you are installing an Extension, as those also block Extension from installing requirements

by. Forge

• For RTX 30 and above, you can add the following flags to slightly increase the performance; but in rare occurrences, they may cause OutOfMemory errors or even crash the WebUI; and in certain configurations, they may even lower the speed instead
  • --cuda-malloc
  • --cuda-stream
  • --pin-shared-memory

by. Classic

• --uv: Replace the python -m pip calls with uv pip to massively speed up package installation
  • requires uv to be installed first (see Installation)
• --uv-symlink: Same as above; but additionally pass --link-mode symlink to the commands
  • significantly reduces installation size (~7 GB to ~100 MB)

Important

Using symlink means it will directly access the packages from the cache folders; refrain from clearing the cache when setting this option

• --model-ref: Points to a central models folder that contains all your models
  • said folder should contain subfolders like Stable-diffusion, Lora, VAE, ESRGAN, etc.

Important

This simply replaces the models folder, rather than adding on top of it

• --persistent-patches: Enable the persistent LoRA patching

  • no longer apply LoRA every single generation, if the weight is unchanged
  • save around 1 second per generation when using LoRA

• --sage: Install the sageattention package to speed up generation

  • will also attempt to install triton automatically

• --flash: Install the flash_attn package to speed up generation

• --onnxruntime-gpu: Install the onnxruntime with the latest GPU support

• --fast-fp16: Enable the allow_fp16_accumulation option

with SageAttention 2

• --sage2-function: Select the function used by SageAttention 2
  • options:
    • auto (default)
    • fp16_triton
    • fp16_cuda
    • fp8_cuda

If you are getting NaN errors, try play around with them

Installation

0. Install git

1. Clone the Repo

   git clone https://github.com/Haoming02/sd-webui-forge-classic

2. Setup Python

Recommended Method

• Install uv
• Set up venv

  cd sd-webui-forge-classic
  uv venv venv --python 3.11 --seed

• Add the --uv flag to webui-user.bat

Deprecated Method

• Install Python 3.11.9
  • Remember to enable Add Python to PATH

3. (Optional) Configure Commandline
4. Launch the WebUI via webui-user.bat
5. During the first launch, it will automatically install all the requirements
6. Once the installation is finished, the WebUI will start in a browser automatically

Attention Functions

Important

The --xformers, --flash, and --sage args are only responsible for installing the packages, not whether its respective attention is used (this also means you can remove them once the packages are successfully installed)

Forge Classic tries to import the packages and automatically choose the first available attention function in the following order:

1. SageAttention
2. FlashAttention
3. xformers
4. PyTorch
5. Basic

Tip

To skip a specific attention, add the respective disable arg such as --disable-sage

Note

The VAE only checks for xformers, so --xformers is still recommended even if you already have --sage

In my experience, the speed of each attention function for SDXL is ranked in the following order:

• SageAttention ≥ FlashAttention > xformers > PyTorch >> Basic

Important

When using SageAttention 2, both positive prompts and negative prompts are required; omitting negative prompts may cause NaN issues

Tip

Check out the Wiki~

Issues & Requests

• Issues about removed features will simply be ignored
• Issues regarding installation will be ignored if it's obviously user-error
• Feature Request not related to performance or optimization will simply be ignored
  • For cutting edge features, check out the Neo branch instead
  • Non-Windows platforms will not be officially supported, as I cannot verify nor maintain them

---

Special thanks to AUTOMATIC1111, lllyasviel, and comfyanonymous, kijai,
along with the rest of the contributors,
for their invaluable efforts in the open-source image generation community