Uploaded September 2026 | Updated November 2026, 1 day ago
I tested the new 3-step LoRA for MiniMax H3 in ComfyUI on one RTX 5090. It is faster, but not 11 times faster, and it has a clear quality cost in some scenes.
Alibaba's TaoLive AIGC team says this LoRA is 11.45 times faster. That number comes from eight H20 GPUs. On my machine, I compared it with the 4-step LoRA that most of us already use. The 3-step LoRA finished in 3 minutes 11 seconds, and the 4-step LoRA took 3 minutes 46 seconds. That is 35 seconds saved, about 15% faster.
In this video, I test three harder scenes:
- A dim room with two people talking, with four shots and hard cuts
- A car braking and sliding on a snowy mountain road
- A woman jumping from a roof and attacking
You will see where the 3-step LoRA works well, where it changes the lighting or breaks the motion, and how a hybrid workflow fixes the motion problem.
Important: right now, this LoRA supports only the FL2VA version of MiniMax H3 (you give a first frame, and it makes the video and audio). I used ComfyUI Kitchen for attention.
Download the 3-step LoRA
There are three converted versions for ComfyUI. Pick one and save it in your ComfyUI/models/loras folder.
- About 2.48 GB:
huggingface.co/Robert1212star/TaoMate-H3-3Step-ComfyUI
- About 1.2 GB:
huggingface.co/Asirus/TaoMate_H3_3_Step_LoRA/blob/main/taomate_h3_3step_MAXQUALITY_bf16.safetensors
- Kijai's rank 19 version, about 182 MB:
huggingface.co/Kijai/MiniMax-H3_comfy/blob/main/loras/minimax_h3_taomate_3step_lora_avg_rank_19_bf16.safetensors
Nobody has shown yet that the small file gives the same quality as the bigger ones. I will test that in a future video.
Original LoRA by the TaoLive AIGC team:
huggingface.co/TaoLiveAIGC/TaoMate-H3
MiniMax H3 base model:
huggingface.co/MiniMaxAI/MiniMax-H3
Workflow:
patreon.com/ComfyuiWorkflowBlog/posts/minimax-h3-3-my-169797848
My answer
Use the 3-step LoRA for still shots, close-ups, and scenes with good light and small camera changes. It also works well as the last stage of a hybrid workflow. For action scenes and dim light, I still use 4 steps.
Have you tried the 3-step LoRA yet? Tell me what you found in the comments.
#MiniMaxH3 #ComfyUI #AIVideo
I tested the new 3-step LoRA for MiniMax H3 in ComfyUI on one RTX 5090. It is faster, but not 11 times faster, and it has a clear quality cost in some scenes.
Alibaba's TaoLive AIGC team says this LoRA is 11.45 times faster. That number comes from eight H20 GPUs. On my machine, I compared it with the 4-step LoRA that most of us already use. The 3-step LoRA finished in 3 minutes 11 seconds, and the 4-step LoRA took 3 minutes 46 seconds. That is 35 seconds saved, about 15% faster.
In this video, I test three harder scenes:
- A dim room with two people talking, with four shots and hard cuts
- A car braking and sliding on a snowy mountain road
- A woman jumping from a roof and attacking
You will see where the 3-step LoRA works well, where it changes the lighting or breaks the motion, and how a hybrid workflow fixes the motion problem.
Important: right now, this LoRA supports only the FL2VA version of MiniMax H3 (you give a first frame, and it makes the video and audio). I used ComfyUI Kitchen for attention.
Download the 3-step LoRA
There are three converted versions for ComfyUI. Pick one and save it in your ComfyUI/models/loras folder.
- About 2.48 GB:
huggingface.co/Robert1212star/TaoMate-H3-3Step-ComfyUI
- About 1.2 GB:
huggingface.co/Asirus/TaoMate_H3_3_Step_LoRA/blob/main/taomate_h3_3step_MAXQUALITY_bf16.safetensors
- Kijai's rank 19 version, about 182 MB:
huggingface.co/Kijai/MiniMax-H3_comfy/blob/main/loras/minimax_h3_taomate_3step_lora_avg_rank_19_bf16.safetensors
Nobody has shown yet that the small file gives the same quality as the bigger ones. I will test that in a future video.
Original LoRA by the TaoLive AIGC team:
huggingface.co/TaoLiveAIGC/TaoMate-H3
MiniMax H3 base model:
huggingface.co/MiniMaxAI/MiniMax-H3
Workflow:
patreon.com/ComfyuiWorkflowBlog/posts/minimax-h3-3-my-169797848
My answer
Use the 3-step LoRA for still shots, close-ups, and scenes with good light and small camera changes. It also works well as the last stage of a hybrid workflow. For action scenes and dim light, I still use 4 steps.
Have you tried the 3-step LoRA yet? Tell me what you found in the comments.
#MiniMaxH3 #ComfyUI #AIVideo










