GGUF is outdated in the latest versions of Comfy-UI. If you want a good balance of size, speed and quality you should use the int8_convrot model from the official Comfy Org Repo https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/diffus...
reply> a ~9-second 480x864 clip at 20 steps takes me a bit over an hour
replythat's rough. for comparison, i tried the exact same parameters on my 5090 RTX and it took 2 minutes to generate.
i believe diffusion models are primarily compute bound so the macs aren't really the ideal hardware for this kind of stuff
This implementation is much faster on my M5 Max, like a few minutes for the same video, but on an M5 Max with 128GB, didn't test on M5 Pro. About memory, could be executed on 64GB with a few changes.
replyMemory bandwidtih between pro and max is double. 300gb/s vs. 600gb/s btw.
replyDoes not matter much in this case. GPU bound.
replyI wonder how much faster your m5 pro is compared to my M1 Max @ 64gb
replywait till the M7 bro you’re almost there, rumor has it that Apple is skipping the M6 but it still might be 2028
replyDiscussion: https://news.ycombinator.com/item?id=48676795
replyWhat is the quality of the output like compared to something like Veo?
replyHow much free space do you have left after running this llm model? Have you tried to develop your own model with the M5?
reply