ComfyUI on AMD ROCm: Local Image Generation Without NVIDIA

ComfyUI runs well on AMD ROCm. The trick to coexisting with a big LLM on one card is aggressive model offloading.

Launch flags that keep VRAM low

Run ComfyUI with --cache-none so models are offloaded between nodes instead of kept resident — peak VRAM drops dramatically, which is what lets image generation share a GPU with a large language model:

python main.py --listen 0.0.0.0 --port 8188 --cache-none

Same GPU passthrough as llama.cpp

devices: ["/dev/kfd", "/dev/dri"]
group_add: ["<render-gid>", "<video-gid>"]
environment: ["HIP_VISIBLE_DEVICES=0"]

Driving it from chat

In ai-stack, Open WebUI calls ComfyUI as a tool: it first unloads the LLM to free VRAM, submits the image job, then frees ComfyUI's VRAM so chat can resume. You get text-to-image and vision from the same chat box, all local.

Image models (diffusion transformer, text encoders, VAE) are large — keep them on fast storage and out of your git repo; fetch them with a checksum-verifying downloader.

FAQ

Does ComfyUI work on AMD without CUDA?

Yes, ComfyUI runs on AMD ROCm. No NVIDIA GPU or CUDA is required.

How do I stop ComfyUI from hogging all my VRAM?

Launch with --cache-none to offload models between nodes, keeping peak VRAM low enough to share the card with an LLM.

Can I generate images from a chat interface?

Yes — Open WebUI can call ComfyUI as a tool, handing VRAM back and forth with the LLM automatically.

Get the full stack on GitHub →