Repository navigation
tutorials/image/minimax/minimax-h3-image #2338
Replies: 2 comments
|
Why is this workflow (text to image) so slow? Literally taking hours and it's not done yet. |
|
Hi @pinasterj-lang, hours is not normal for this model, so the cause is almost certainly memory, not the workflow. At 4 MP a run on one recent GPU is minutes. The weights alone are 46 GB before sampling starts: a 33.1B diffusion transformer (21 GB, INT8) plus the Qwen3-VL-32B text encoder (25.28 GB, INT8). On a GPU with less VRAM than that, ComfyUI keeps part of each model loaded and streams the rest from system RAM while sampling, which turns each step into minutes. The text encoder is often the real cost, even at a small canvas. Check the ComfyUI console when the run starts:
Two settings are also worth checking:
As a sanity check, run the template unchanged at 4 MP and see whether it finishes in minutes. That separates a VRAM problem from a broken install. If it still crawls, send your GPU model and VRAM, OS, ComfyUI version, and the first 30 lines of console output from a run, and I can point at the exact cause. There is also an open report of full-resolution runs on the H3 family getting much slower on recent builds (Comfy-Org/ComfyUI#15665), so if you are on 0.32.0 or newer at high megapixels, that is worth comparing against. Page: https://docs.comfy.org/tutorials/image/minimax/minimax-h3-image |
Uh oh!
There was an error while loading. Please reload this page.
tutorials/image/minimax/minimax-h3-image
Run MiniMax-H3-Image in ComfyUI: one open-weights checkpoint for text to image and prompt-driven image editing at about 16 MP, distilled to 6 sampler steps.
https://docs.comfy.org/tutorials/image/minimax/minimax-h3-image
All reactions