AMD RX6800 - Really slow at Diffusion models specifically?
retained text
So just wanted to check if anyone knows more than I do (not a high bar to clear to be honest), because I am noticing a recurring issue when using Comfyui ROCM. Diffusion models (and Video generation models) are insufferably slow. I am basically stuck using nothing but either Noob or Illustrious models for generations because all the other options (Krea, Anima, MOSS, Flux, Chroma, etc) are significantly slower. I've had this issue for over a year. What I mean as an example, generating a 1280\*720 (\~1 megapixel) resolution file: * 6.6gb Illustrious XL model = **1.2it/s** \* 50steps = **75 seconds** * 5.04gb MOSS model = **9.66s/it** \* 50steps = **490 seconds** * 9.46gb MOSS model = **13s/it** *(fair enough, it's bigger) =* **653 seconds** * 3.89gb Anima model = **2.2s/it** \* 50 steps = **109 seconds** Like... why are these smaller models so horrendously slow, normally one would think the larger model would be slower? I can tell my Vram is fine, as I still have at minimum 2gb headroom even on the largest MOSS model while running. Even the Anima model which is almost half the size of Illustrious takes 50% longer for a single image. The gap gets bigger when upscaling, exponentially so, up to 10-15x longer. I can legitimately generate and upscale five images to 4K in Illustrious XL (takes under 2 minutes each when I run in the normal step count) in the time it takes to make a single low resolution image with MOSS for example. I tested with smaller text encoders, smaller VAE's as well - they make literally 0 difference in speed but do have a noticeable loss in quality, so I just keep those on the original versions. And Video generation is even worse: * LTX 2.5 Q-4\_k\_M (13.2gb), 24fps for 5 seconds at 738\*416 (\~0.3 megapixel): * 70s/it for 8 steps, followed by * 15-20 minutes of pure decoding (tiled, 512) It takes a bit over 25 minutes to run the full size LTX model, so basically no difference in generation time compared to the GGUF. At least in this case I can see it is offloading a substantial amount to RAM so the full size model's time makes sense but I don't get why the GGUF version is seemingly running significantly slower than it should be. I can keep the entire GGUF model loaded in VRAM after closing some background apps, it's only when it starts to decode that it starts using about 40gb of system memory - what would a realistic generation time for this be given my specs? Specs: \- Latest Comfyui ROCM Version (I updated today, also updated ROCM) \- AMD RX 6800 16GB VRAM \- 80gb System memory \- Ryzen 5 5600 CPU \- Windows 11 Any help, or even any insight anyone could give would be greatly appreciated.
document fetched 2026-10-21 00:41 · canonical https://reddit.com/r/comfyui/comments/1xaup2o/amd_rx6800_really_slow_at_diffusion_models