Innovate Futures @ Benji

Minimax H3 - Making Combat , Fast Motion Scenes With Local AI (Patreon)

Published:

2026-09-08 13:01:22

Edited:

2026-09-08 22:18:40

Imported:

2026-09

Videos

vid-vfi_00013-audio.mp4

Downloads

Content

https://youtu.be/uhG307gPHco

Hey everyone.

New video is up, and this one is for anyone who has watched those crazy Seedance fight scene demos and wondered if we can do the same thing locally.

Short answer: yes. In ComfyUI, with MiniMax H3, I'm getting high-density martial arts combat — sword fights, kinetic energy explosions, characters getting hit and reacting properly — all generated on my own machine.

The video tests two ways of doing it. First, a stack of LoRAs: a wushu action LoRA for the martial arts motion, a combat physics LoRA for the impacts, and a motion repair LoRA so nobody's legs snap in half mid-jump. Second, a fully fine-tuned diffusion model called Singularity that does it all in one model, no trigger words, no stacking.

One thing that surprised me: the LoRA stack is genuinely good, but the fine-tuned model pulled ahead in the side-by-side. The hits land harder, the reactions are cleaner, and it doesn't get that pixelated dust in the air that the LoRA runs sometimes have. I even prompted the dust on purpose to see if it would break — it didn't.

I also stitched the generated shots into a little warehouse story, which honestly might be my favorite part.

Links to everything are in the YouTube description. Thanks for watching, and I hope you have as much fun with this one as I did.

LoRAs

https://huggingface.co/Jojocodex/wushu-action-v7-minimax-h3-fl2va-ref2va-lora/

https://huggingface.co/JOKER141/MiniMax-H3-Combat-Base-V2

https://huggingface.co/JOKER141/MiniMax-H3-General-Motion-Continuity-Repair

Diffusion Model - Fine Tune

https://huggingface.co/WarmBloodAban/Minimax-h3_Singularity

2 DLSS5 Nodes for options:

https://github.com/Blueforcer/ComfyUI-DLSS5-Enhancer#installation

https://github.com/Konohamaru04/ComfyUI-NVIDIA-DLSS-Frame-Interpolation

update:

Patreon Supporters , your additional content : https://www.patreon.com/aifuturetech/posts/h3-combat-scenes-168955413/

— Benji

Files

Minimax H3 - Making Combat , Fast Motion Scenes With Local AI

Two ways to create high-density martial arts fight scenes locally in ComfyUI with MiniMax H3: a stack of three LoRAs (JoJoCodex Wushu Action v7 at 2,000 steps, Combat Base V2 for impact physics, and General Motion Continuity Repair), versus the Singularity fully fine-tuned diffusion model that supports text-to-video, image-to-video, reference-to-video, and video-to-video in one model. Full workflow walkthrough: trigger words, dual-stage sampling, Turbo LoRA, VDN-H3, DLSS 5 enhancement, and side-by-side comparisons. ComfyUI users and AI video creators who want fighting-game style combat scenes (Tekken/MK grammar, wushu sword fights, barehand duels) generated locally instead of paying for cloud services like Seedance 2.0/2.5. Also for anyone on low-VRAM setups, since the pipeline covers chunked feeding and practical sampling settings. Seedance-quality action scenes have been cloud-only. This video shows the fight scenes are now reproducible locally in ComfyUI, and answers the real question: is a LoRA stack or a fully fine-tuned model the better base for combat motion? The comparison shows the fine-tuned Singularity model wins on impact physics, hit reactions, and texture cleanup, with no trigger words needed. Freebie Workflow https://www.patreon.com/aifuturetech/posts/minimax-h3-fast-168950662?utm_source=youtube&utm_medium=video&utm_campaign=202609 For Patreon Supporters https://www.patreon.com/aifuturetech/posts/h3-combat-scenes-168955413/ LoRAs https://huggingface.co/Jojocodex/wushu-action-v7-minimax-h3-fl2va-ref2va-lora/ https://huggingface.co/JOKER141/MiniMax-H3-Combat-Base-V2 https://huggingface.co/JOKER141/MiniMax-H3-General-Motion-Continuity-Repair Or Diffusion Model - Fine Tune https://huggingface.co/WarmBloodAban/Minimax-h3_Singularity Timeline: 00:00 Introduction 00:15 Stacked LoRAs vs Singularity 01:30 LoRA Stack Breakdown 03:25 Singularity Model Features 03:46 Workflow Setup 04:12 LoRA Node Settings 05:18 Sampling & DLSS 5 Upscaling 06:46 Text-to-Video Test 08:02 Image-to-Video Test 09:06 Singularity Model Test 10:05 Model Comparison 11:08 Reference-to-Video Demo 12:18 Storytelling & Scene Sequence 13:30 Optimization Tips Local Workstation GPU : https://amzn.to/3XfXsAO -------------------------------------------------------------------------------------------------------------------------------- If You Like tutorial like this, You Can Support Our Work In Patreon: https://www.patreon.com/c/aifuturetech #comfyui #minimaxh3 #localai #aivideo