It seems the LoRA for multi-person videos isn't working.
First, I’d like to thank KJ for contributing this acceleration LoRA to the community. I’ve been using minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy.safetensors locally; it significantly speeds up single-person video generation—cutting the time in half or even more. However, after generating several videos, I noticed that the LoRA seems to fail when generating scenes with multiple people, resulting in heavily pixelated (mosaic-like) visuals. I tested various prompts for multi-person videos, and they all produced this pixelated output. I also experimented with different samplers (such as er_sde, sa_solver, and res_multistep) and step counts (like 4 or 8 steps), but the issue persisted. If a formal version of the LoRA is released in the future, please look into this problem.
It's probably related to the total token count, which is a total of resolution, duration and references used, there are some issues in sageattention itself that make it overflow after certain points, fully corrupting the output. I have fix for the ~100k tokens case in KJNodes sageattention + memory efficient sage attention nodes, if it still overflows with those we may need to figure out where else it can overflow.
It's probably related to the total token count, which is a total of resolution, duration and references used, there are some issues in sageattention itself that make it overflow after certain points, fully corrupting the output. I have fix for the ~100k tokens case in KJNodes sageattention + memory efficient sage attention nodes, if it still overflows with those we may need to figure out where else it can overflow.
Now I see the problem why 2K upscale workflow can't exceed 5-7 seconds, you offered a plausible explanation.
works great for me kaji has the best models
can u make the 4 step turbo of this minimax_h3_ref_lora_rank_256_bf16.safetensers please ?
It's probably related to the total token count, which is a total of resolution, duration and references used, there are some issues in sageattention itself that make it overflow after certain points, fully corrupting the output. I have fix for the ~100k tokens case in KJNodes sageattention + memory efficient sage attention nodes, if it still overflows with those we may need to figure out where else it can overflow.
I tried generation without applying sage-attention, now I can generate 15s upscaled 2k video. I'm using the latest version of sage attention. Do I need to upgrade pytorch as well?