Great work! Please also make one for the ref2va version. Thanks!
Great work!! Please Kijai, also make one for the ref2va version, that would be great! Thank you very much for your efforts for the open-source community. You are the best!
One question... Do you plan to make a 4-step turbo version for the minimax h3? That would be fantastic, it would be a huge help because most users have low/medium VRAM.
Will do that version once the implementation is merged and verified.
I don't train the distill loras, that's gonna be up to teams like Lightx2v or more experienced trainers like Ostris etc.
Great! Thank you again.
I understand. Yes, I just saw that Ostris is working on a 4 steps lora and the results are simply incredible! This will be a game-changer for everyone. ;)
Great work!! Please Kijai, also make one for the ref2va version, that would be great! Thank you very much for your efforts for the open-source community. You are the best!
One question... Do you plan to make a 4-step turbo version for the minimax h3? That would be fantastic, it would be a huge help because most users have low/medium VRAM.
I think distillation or 4step LoRA will turn this beautiful model into a glorified 40GB slideshow maker. have you tried 8 steps with dpmpp_2m and sgm_uniform?
On my hardware (RTX 5080), it generates 8 seconds clips (480×864) faster than wan2.2 4step (480×832 - 5 seconds)
thx your hard work, kj, ref2va verson plz
Great work!! Please Kijai, also make one for the ref2va version, that would be great! Thank you very much for your efforts for the open-source community. You are the best!
One question... Do you plan to make a 4-step turbo version for the minimax h3? That would be fantastic, it would be a huge help because most users have low/medium VRAM.
I think distillation or 4step LoRA will turn this beautiful model into a glorified 40GB slideshow maker. have you tried 8 steps with dpmpp_2m and sgm_uniform?
On my hardware (RTX 5080), it generates 8 seconds clips (480×864) faster than wan2.2 4step (480×832 - 5 seconds)
the current issue is comfyui's poor memory management, not inference itself. it keeps loading models in and out, to ssd, and it's overall just not great. inference is only an issue on GPUs below 3080 where it basically takes 2x as long
I think distillation or 4step LoRA will turn this beautiful model into a glorified 40GB slideshow maker. have you tried 8 steps with dpmpp_2m and sgm_uniform?
On my hardware (RTX 5080), it generates 8 seconds clips (480×864) faster than wan2.2 4step (480×832 - 5 seconds)the current issue is comfyui's poor memory management, not inference itself. it keeps loading models in and out, to ssd, and it's overall just not great. inference is only an issue on GPUs below 3080 where it basically takes 2x as long
I had the same issue with ComfyUI dynamic memory management on an RTX 3080 (64 GB DDR4, Samsung SSD). I replaced the SSD with a 1 TB M2, and everything changed for the better.
Another issue I ran into was after upgrading to an RTX 5080 and 96 GB of RAM I switched to Linux to reduce overhead, but I couldn’t use dynamic memory management.
After 2 or 3 runs it would always hit OOM, and I had to restart the PC manually. It turned out to be a bug in Wayland with Nvidia GPUs.
Now it’s solved, and I can say wholeheartedly that ComfyUI dynamic memory management is a state of the art solution.
I can run full Wan 2.2 FP32 and UMT5_xxl FP32 with 3 gaming monitors running at 120 Hz, plus a browser open with at least 3 or 4 tabs. I mean when OS and hardware are in harmony, ComfyUI really shines.