Stable Diffusion WebUI Codex

Local-first FastAPI + Vue 3 WebUI for running modern diffusion pipelines with managed installers and safe updates.

Quick links: Install | Quick Start | Docker Install | Model Hub Notes | Custom PyTorch FA2 | Support

Preview screenshots

Launcher (Services tab)	WebUI (SDXL tab)
Manage API/UI services and runtime startup state from one place.	Run SDXL workflows with generation controls and model/runtime options.

Custom PyTorch builds (FlashAttention2)

Important

This project optionally provides Windows PyTorch builds with FlashAttention2 enabled for RTX architecture targets.

• Repository: https://github.com/sangoi-exe/pytorch
• Targeted RTX build variants:
  • SM80
  • SM86
  • SM89
  • SM90

For this workflow, these custom builds are the recommended path when you want FA2-enabled runtime behavior.

What you can run

• Multi-engine workflows: SD15, SDXL, FLUX.1, Z-Image, WAN 2.2, and related adapters.
• Utility views and tools: PNG Info, XYZ plot, GGUF tools, and workflow snapshots.
• Local-first backend/UI stack with managed Python and Node environments.

Quick start

Prerequisites:

• Git
• Internet access for first install (uv, managed Python, repo-local Node.js via nodeenv, and Python wheels)

For detailed install options, backend selection, and troubleshooting, see INSTALL.md.

Windows

.\install-webui.cmd

install-webui.cmd
run-webui.bat

Online one-liner + full manual (no installer scripts) are documented in INSTALL.md.

Optional safe update:

update-webui.bat

Linux / WSL

bash install-webui.sh
./run-webui.sh

Optional safe update:

bash update-webui.sh
# Optional: ignore untracked-path preflight checks (tracked changes still abort)
bash update-webui.sh --force

On Linux/WSL, startup prints the local URLs:

[webui] API: http://localhost:7850
[webui]  UI: http://localhost:7860

Open the UI URL in your browser. Stop with Ctrl+C.

On Windows, use the launcher output/UI to confirm the active API and UI endpoints.

Docker Installation Guide

Use this path when you want containerized setup with persisted models/output state.

1) Build the image

docker build -t codex-webui:latest .

Optional build overrides:

docker build -t codex-webui:latest . \
  --build-arg CODEX_TORCH_MODE=cpu \
  --build-arg CODEX_TORCH_BACKEND=cu126

2) Run with GPU and persistent volumes

docker run --rm -it --gpus all \
  -p 7850:7850 -p 7860:7860 \
  -v "$(pwd)/models:/opt/stable-diffusion-webui-codex/models" \
  -v "$(pwd)/output:/opt/stable-diffusion-webui-codex/output" \
  codex-webui:latest

3) First-time profile setup only (no API/UI start)

docker run --rm -it \
  codex-webui:latest --tui --configure-only

4) Compose path (non-interactive default)

docker compose up --build

First-time interactive profile setup with compose:

docker compose run --rm webui --tui --configure-only

Docker runtime notes

• Entrypoint is run-webui-docker.sh (delegates to apps/docker_tui_launcher.py and then run-webui.sh).
• Disable interactive TUI with -e CODEX_DOCKER_TUI=0 or runtime args --no-tui --non-interactive.
• If you override API_PORT_OVERRIDE or WEB_PORT in TUI/profile, host -p mappings must use the same ports.
• Docker defaults are preseeded for this project profile (CUDA runtime path, SDPA flash, LoRA online, WAN22 ram+hd) and can be overridden with -e KEY=VALUE or compose .env.
• Default allocator env is PYTORCH_CUDA_ALLOC_CONF=backend:cudaMallocAsync.

Safe updater behavior

update-webui.(bat|sh) is fail-closed and non-destructive:

• Dirty worktree checks are strict: tracked changes always abort; untracked changes abort unless --force is set.
• --force only relaxes untracked-path preflight checks; git pull --ff-only safety checks still apply.
• Ignored paths (.gitignore) are excluded from dirty-tree abort checks.
• Update path remains non-destructive: git fetch --prune + git pull --ff-only.

Full contract and edge cases are documented in INSTALL.md.

Models

Official Hugging Face repo: https://huggingface.co/sangoi-exe/sd-webui-codex

Highlights from the Hub (snapshot from hf download --dry-run on 2026-01-29):

• FLUX.1
  • flux/FLUX.1-dev-Q5_K_M-Codex.gguf
• Z-Image
  • zimage/Z-Image-Turbo-Q5_K_M-Codex.gguf
  • zimage/Z-Image-Turbo-Q8_0-Codex.gguf
• WAN 2.2
  • wan22/wan22_i2v_14b_HN_lx2v_4step-Q4_K_M-Codex.gguf
  • wan22/wan22_i2v_14b_LN_lx2v_4step-Q4_K_M-Codex.gguf
  • wan22-loras/wan22_i2v_14b_HN_lx2v_4step_lora_rank64_1022.safetensors
  • wan22-loras/wan22_i2v_14b_LN_lx2v_4step_lora_rank64_1022.safetensors
  • wan22-loras/wan22_t2v_14b_HN_lx2v_4step_lora_rank64_1017.safetensors
  • wan22-loras/wan22_t2v_14b_LN_lx2v_4step_lora_rank64_1017.safetensors

Default model roots are under models/:

• models/sd15/, models/sdxl/, models/flux/, models/zimage/, models/wan22/
• plus *-vae, *-tenc, *-loras

More model hub and folder layout details: README_HF_MODELS.md. If you customize model roots, edit apps/paths.json.

Support

• Bug reports: https://github.com/sangoi-exe/stable-diffusion-webui-codex/issues
• Questions and discussion: https://github.com/sangoi-exe/stable-diffusion-webui-codex/discussions

License (noncommercial)

• Code license: PolyForm Noncommercial License 1.0.0 (LICENSE).
• Required notice must be preserved: NOTICE.
• Commercial use is not permitted: COMMERCIAL.md.
• Trademarks and branding: TRADEMARKS.md.

Credits

• Diffusion ecosystem baseline: Hugging Face Diffusers.
• UX and workflow inspiration: AUTOMATIC1111 and Forge.