ComfyUI on AMD: Windows ROCm setup and GGUF comparison

This tutorial sets up a ComfyUI ROCm fork for local AI image generation on Windows. The speaker describes support for RDNA 1 through 4 and cites the repository author's RX 6800 XT test. These are reported compatibility claims; the demonstration does not test every AMD GPU.

The setup starts in PowerShell with Git, a repository clone and install.bat. Prerequisites include a runtime library and Visual Studio Build Tools. The installer bundles Python and detects the GPU architecture. The speaker checks that PyTorch reports ROCm as available before continuing. ComfyUI-user.bat launches the server, after which the tutorial checks the selected GPU and opens the browser interface.

The first Z Image Turbo workflow uses an approximately 11 GB diffusion model with a Qwen 3 4B text encoder and a VAE. Its first image takes 98 seconds, and a repeat run does not improve speed. The tutorial then installs ComfyUI-GGUF through the manager and switches to Q4 models, reducing the diffusion model to about 4.5 GB.

For the comparison, the speaker matches the prompt, sets the output to 1024 × 1024 and uses four sampling steps. The quantized workflow takes 32 seconds initially and 19.85 seconds with models already loaded. Those timings describe this setup, rather than a general AMD benchmark. Text rendering remains inconsistent: the faster repeat run merges the requested words into an incorrect result.