Draft Your AI Fight Scene Before You Render It: A Two-Pass MiniMax H3 Workflow
Fight scenes are where AI video falls apart. Fast limbs smear, weapons pass through bodies, and the hero who started the scene with a halberd ends it holding a sword. Every video model struggles with fast, continuous action, and Kling's own launch material this week names fast movement and continuous action as the areas where models are under "the most pressure."
So the interesting part of a new fight-scene workflow posted to Hugging Face on September 30 is not that it makes fights. It is how it handles the fact that most attempts will fail.
The workflow, published by Work-Fisher for the open MiniMax H3 video model, shows you a rough draft of the fight halfway through. If the choreography is wrong, you cancel. You only spend the expensive part of the render on takes that already move correctly.
That habit, draft then commit, is worth stealing even if you never make a martial-arts video.
What the workflow does
The MiniMax H3 Wuxia Fight Workflow is a ComfyUI graph, the node-based visual editor many people use to run open image and video models on their own machines. You give it one reference picture of your character and a written description of the fight. It returns up to 15 seconds of continuous action, one unbroken take, with sound.
It works in two passes:
- Pass one renders the whole fight small and rough. This is where the model decides the motion, the camera path and the composition.
- An upscaler enlarges that rough version by one and a half times.
- Pass two goes back over the enlarged version and adds the fine detail: faces, fabric, sparks, debris.
Between pass one and pass two, a preview window shows you the draft. That is the whole trick.
The author tested it on an RTX 4070 Ti, a mid-range card with 12 GB of memory, with 32 GB of system memory. That puts it within reach of a lot of home setups, not just rented cloud machines.
Why the draft pass changes the economics
Here is the mechanism in plain terms. When a video model works, the big decisions happen early and cheaply: where the bodies go, which way the camera moves, roughly what each moment looks like. The expensive work is detail at full size. If the punch in your draft lands in the wrong place, no amount of detail will fix it. You are polishing a mistake.
Most single-pass workflows make you wait for the polished version before you find out the motion was wrong. This one shows you the motion first.
For fight scenes this is close to essential. Action is the content most likely to fail, so it is also the content where catching failure early saves the most. I would rather throw away five rough drafts in the time it takes to finish one full render than finish five full renders and keep one.
The author's own advice points the same way: start with roughly 5-second clips, not the full 15. Short drafts fail faster, and you learn what your prompt is doing before you commit to a long take.
The prompt template is the second lesson
The download includes a Chinese-language prompt template, 打戏提示词模版_MiniMax15秒一镜到底_v1.4.md, which roughly means "fight-scene prompt template, MiniMax, 15 seconds, one continuous take." It is the most useful thing in the folder for anyone who writes action, in any tool.
Its core idea: a fight is a timeline, not a mood. The template splits 15 seconds into eight beats. The first five run 1.5 seconds each, the sixth runs 2 seconds, the seventh 2.5, and the finale gets 3. Each beat holds one exchange: an attack, a response, and a counter or result, with no more than three or four actions in a row.
Several rules in it apply to any action prompt:
- Give the fight an arc. The template lays out an opening probe, a small early advantage for the hero, the opponent dominating twice, an escalation, a full clash and a narrow win. Without one, a fight is just motion.
- Lock each character's state. If the hero is on horseback, they stay on horseback, and the big finishing power move fires only once. Change as little as possible from beat to beat. This stops the model from drifting, the reason swords become halberds.
- Fix the camera. One continuous handheld shot with a slight shake. No cuts, no slow motion, no static camera, with a single freeze saved for the very end.
- Every hit leaves evidence. Dust, sparks, cracks or debris on every beat. The template bans blood and suggests fire, frost particles or armor fragments instead.
The worked example in the template uses that last rule cleverly. A crack opens in the opponent's sword handle in beat three, and the final strike in beat eight drives through that same crack. A detail planted early pays off at the end, which is how a fight choreographer thinks.
Put this into practice
1. Check your machine. You need an NVIDIA card (tested on 12 GB), 32 GB of system memory, ComfyUI 0.36.0 or newer, and about 44 GB of free disk space for the model files.
2. Install the six add-on node packs at the versions listed on the card. The core one is comfyui-minimax-h3-audio-T8 at version 1.86.0. The others are KJNodes, VideoHelperSuite, Comfyroll, utils-nodes and LayerStyle. Match the versions. Mismatched versions are the most likely reason the workflow will not load.
3. Download the twelve model files into the folders the card names. One upscaler file needs renaming after download; the card gives the exact names. Do this carefully, because a misplaced file usually fails with an unhelpful error.
4. Load the workflow file (【WORK-FISHER】26-9-28-MINIMAXH3-武戏高动态工作流.json) into ComfyUI.
5. Add your reference picture in the group labelled "参考图1" (Reference Image 1). The interface labels are in Chinese, so a translation app on your phone helps the first time.
6. Write your fight in the box labelled "简单提示词写入" (simple prompt). Use the eight-beat structure from the template. The template itself is written in Chinese and tells you to write the final prompt in Chinese, and the card never says English works. Translate your beats into Chinese for the closest match to the author's results, or test English yourself and compare.
7. Set the duration to about 5 seconds for your first runs, and queue.
8. Watch the pass-one preview. Ask one question: does the motion work? Ignore blur and missing detail; pass two handles those. If a limb goes the wrong way or a character swaps weapons, cancel and adjust the prompt.
9. Let the good drafts finish. Output lands in ComfyUI's output folder.
10. Scale up. Once your 5-second beats land reliably, move to 15 seconds using the full eight-beat timeline.
If you do not care about fights, the transferable step is 8. Look for a preview of any workflow you run, and get into the habit of judging motion before detail.
Honest limitations
The download is huge. Twelve files, 43.9 GB. The main video model alone is about 19.5 GB and the text reader about 14.6 GB. Budget the time and the disk.
The setup is fragile. Six node packs pinned to specific versions can clash with what you already have installed. Consider a separate ComfyUI install for this.
The interface and template are in Chinese. Very workable with a translator, but slower the first time.
Fifteen seconds is the ceiling. If you need a longer fight, you will be stitching takes, with the continuity problems that brings.
Licenses stack up. MiniMax H3 runs under the MiniMax H3 Community License, and the version of that license quoted on related model cards grants no rights in the EU, the UK, South Korea or the United States. On top of that, the workflow uses a converted base model and six add-ons from several different authors, plus a bridge model, each with its own terms. Read all of them before you use the output in paid work.
The draft does not guarantee the final. Pass two adds detail, and detail can still go wrong: hands, faces, weapon edges. The preview catches bad motion, not every flaw.
No timing figures are published. The card does not say how long each pass takes on the test card, so you will find out on your own hardware.
The habit outlasts the workflow
MiniMax H3 will be replaced. This particular graph, with its pinned versions and Chinese labels, will break the next time one of its node packs updates. The two ideas inside it will last longer.
The first is draft then commit. Spend a little to see the motion, then spend a lot only on what already works. It is how animators work with rough passes, how editors work with offline cuts, and now how local video generation works on a modest card.
The second is that action needs a stopwatch more than adjectives. "An epic battle" gives the model nothing. "1.5 seconds: she parries, sparks, steps back" gives it a job.
Try the template on one short fight this week, even if you write it for a different video tool. Time the beats, lock the states, plant one detail early and pay it off late. Then watch the rough draft and ask whether you would cut it. If the answer is no, you have learned more about directing AI action than any amount of rerolling would teach you.
Medium metadata
Title: Draft Your AI Fight Scene Before You Render It: A Two-Pass MiniMax H3 Workflow
Subtitle: A free ComfyUI workflow shows you a rough version of the fight before the expensive render, and its eight-beat template shows how to write action an AI can follow.
Tags: AI Video, ComfyUI, Action Scenes, Filmmaking, Generative AI
Estimated read time: 8 minutes