sd-forge — SD WebUI Forge
A1111-compatible fork by lllyasviel focused on faster inference and lower VRAM use, with native support for SVD, Flux, and modern attention kernels.
Image tag
docker.io/manvarharsh/sd-forge:cuda12
What’s in this image
- Base:
nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04 - Python 3.10 (conda)
- PyTorch with CUDA 12 support
- lllyasviel/stable-diffusion-webui-forge
- xformers
- OpenSSH server
Default ports
| Port | Service |
|---|---|
| 22 | SSH |
| 7860 | Forge WebUI |
Use cases
- A1111 workflows on lower-VRAM GPUs (e.g. RTX 3060, T4)
- Faster image generation than vanilla A1111
- Flux.1 schnell / dev image generation
- Stable Video Diffusion (SVD) experiments
Environment variables
| Variable | Description |
|---|---|
ENABLE_SSH | Enable SSH server |
ENABLE_FORGE | Start the WebUI on port 7860 |
FORGE_EXTRA_ARGS | Extra CLI args (--xformers, --cuda-malloc) |
SSH_PUBLIC_KEY | Public key for SSH |
Persistence
Mount at /data. Forge follows A1111’s directory conventions — checkpoints under /data/models/Stable-diffusion/, LoRAs under /data/models/Lora/, outputs under /data/output/.