Very long processing ...
Hi! I am very delighted by a Multishot workflow(s) becuse I would need up to 60 sec videos and my pc configuration is (I think) capable to do this with Minimax H3. The only problem is very loooong processing - up to 7 hours (30 sec video complete) on my pc with RTX 2080 Ti with 11 Gb VRAM and 64 GB of system RAMJ. Sage attention is not installed, maybe there is a way to cut down the processinbg time ??
Hey, so you have rather low specs for H3, the models are huge and require lots of VRAM and RAM. I've done my best to quantize models to help this out, so I would suggest using this model:
And this text encoder:
Hopefully those will help cut your time down a lot.
Many thanks! You are great! LTX2.3 leaded me to errors (LTX2.2 worked). H3 does not show errors but it is for now slow (maybe upcoming LOra's will help?. The error problem that I had with H3 was with gguf Q4 that was recomended . That leaded to error and I had to choose original 19 Gb diff. mode. (I did not try that new from today). Anyway, you are doing a great job. The time of about 2 hours to generate 30 sec of video would be great for me.
I have tried, for some reason all that gguf's lead to Error(s) ... :(
The GGUF errors are the same architecture issue, not your hardware: run apply_gguf_arch_patch.py from the ComfyUI-H3-Multishot pack and restart. Separately, on 11 GB you'll want Q4_0 for the DiT plus the smallest encoder you can tolerate, and installing Sage Attention is the single biggest speed win available to you.
many thanks for your time for answer! Yes, now it works! Only I have to find optimal encoder.