MiniMax H3 Martial Arts LoRA: 3 Files, 2 Triggers

I went to download the MiniMax H3 martial arts LoRA expecting one file. The repository has three, and the one with the highest step count turned out not to be the newest version of the other two.

The two main checkpoints answer to different trigger words, want different prompt structures, and were trained on different footage. Load one and write the other one's prompt, and you get either the LoRA's weak default behavior or nothing at all, depending on which one you picked.

What is in the repository

Jojocodex/minimax-h3-wushu-action-lora is a community LoRA trained on Comfy-Org's MiniMax-H3 conversion, published under Apache-2.0. Three checkpoints, 155 MB each. The two wushu_action files are pruned; the KungFu file is not marked as such:

One README covers all three files, in two numbered sections that do not always agree with each other.

The card documents only v5 — v4 appears in the file list and is never mentioned. Treat v5 as the maintained one and v4 as a leftover, and assume the same trigger word until something says otherwise.

The KungFu checkpoint is a separate line. The card says it was trained on separately collected footage, and it carries its own trigger word and its own prompt template. Reading 4500 as "newer and better than 3000" is the easiest mistake to make here, and it is why people load the file and then write a prompt it was never trained to answer.

As of August 20, 2026, the repository shows roughly 684 monthly downloads, 29 likes, and two Spaces, with WaveSpeed listed as a text-to-video inference provider. That is a small community project, not a MiniMax release.

The two checkpoints do different jobs

The clearest way to tell them apart is to read what their prompt vocabulary is actually about.

The wushu_action side is about bodies. Weight transfer, hip and shoulder rotation, explosive leg extension, balanced landings, grounded footwork, a low bow stance. The card is explicit that training captions were written as technique plus body coordination plus force application.

The kungfu combat side is about scenes. Dust and motion blur, misty bamboo forests, neon-lit alleys, handheld camera shake, sweeping crane shots, blockbuster movie style. Same subject matter, completely different vocabulary.

Neither one is better. Pick by which half of the shot has to hold up:

• The movement has to read as real — a kick that carries weight, a punch that lands with the hips behind it. That is wushu_action.
• The scene has to read as cinema — a wuxia duel in a bamboo grove, a gritty street fight under neon. That is kungfu combat.

Writing for wushu_action

The trigger word goes at the very front. Every FIGHT-class caption in the training data started with it, so the model has it bound tightly to martial arts motion. Without it the LoRA still has some action prior, just a much weaker one.

wushu_action, [action subject] + [technique and force]
+ [scene and camera] + [optional quality terms]

A complete example from the card:

wushu_action, a wushu practitioner executing a rapid combination of
punches and kicks, powerful hip rotation and weight transfer, gym
background, full-body shot

The second segment is where this LoRA earns its keep. "Performing a spinning kick" is a label; "a spinning kick, explosive leg extension, balanced landing" describes how the body produces and absorbs the force. The training captions were written the second way, so prompts written the second way land closest to what the model learned.

Three phrases from the card's keyword library do most of the work:

• weight transfer — where the mass goes during the strike
• coordinated hip and shoulder rotation — where the power comes from
• grounded footwork — what stops the whole thing from floating

For two-person work, two fighters exchanging strikes gets you the exchange, and realistic impact and recoil is the phrase that addresses contact specifically. Weapons follow the same shape: name the weapon, then name what it does — fighting with a sword, slashing and thrusting, with precise technique.

Writing for kungfu combat

The card counts this one differently. wushu_action's "4-part" structure excludes the trigger; kungfu combat's list counts the trigger as slot one. Line them up on the same basis and it is four content slots against six:

kungfu combat, [subject] + [action] + [environment] + [camera language]
+ [style and texture] + [quality terms]

Two of those six are new. kungfu combat splits scene and camera into separate slots, and it adds style and texture, which wushu_action has no slot for at all. Style language still fits on the wushu side — the card's own sword example ends dark studio, dramatic lighting — but it has to ride along in the scene segment. Here it gets a slot of its own, and the vocabulary offered for it is much wider: wuxia, gritty, anime, film grain, golden hour, neon lights.

Section 2 of the card also notes that this trigger takes weighting syntax — (kungfu combat:1.2) — to strengthen activation. Section 1 does not mention it for wushu_action, though weighting is a feature of the prompt parser rather than of the checkpoint, so it is worth trying on both.

One more difference in that section: kungfu combat is described as required for the LoRA to activate at all, where wushu_action degrades to a weaker prior without its trigger. Same instruction, different consequence for forgetting it.

kungfu combat, ancient Chinese wuxia style, two warriors in traditional
robes, flying kicks and sword clashes, bamboo forest, misty atmosphere,
slow motion impact, epic music vibe, cinematic composition

Note where the style term actually sits in the card's own flagship example: second, not fifth, and the prompt carries no quality terms at all. The slot list is a checklist of what to include, not an order the examples obey. Note also how little of it is about anatomy — the body-mechanics vocabulary that carries wushu_action is almost entirely absent.

The slow-motion trap

The card's suggested negative prompt for wushu_action reads blurry, distorted, low quality, jittery motion, extra limbs, deformed hands, morphing bodies, inconsistent lighting, flickering, static pose, frozen motion.

Example 8 on the same card asks for slow motion feel on a mid-air strike, and slow motion impact is a positive term in the kungfu combat wuxia example.

That is not sloppiness. Choreographed slow motion is a deliberate cinematic effect; a fight that drifts because the model never committed to the strike is a failure. They look nothing alike on screen and identical in a prompt. So be specific about which one you are asking for, and if you are not deliberately asking for it, keep frozen motion in the negatives.

One more thing about negatives: the card says training ran with an empty negative prompt, so the model never learned negative concepts directly. Treat negatives here as cleanup for artifacts you actually see, not as a preventative block you paste into every run.

Before you run it at all

The LoRA weights are Apache-2.0. That covers the adapter, not the base model underneath it — and the base model is where the restriction lives.

The MiniMax H3 Community License excludes four territories from local use: the United States, the European Union, the United Kingdom, and South Korea. If you are in one of them you are not licensed to run H3's weights locally, which means you are not licensed to run this LoRA locally either — an Apache-2.0 adapter is inert without the base model it adapts. Inside the applicable territory, commercial use is allowed up to $20 million in annual revenue. The open-weights breakdown has the full terms.

None of that turns on whether the work is commercial. Read it before you download 155 MB you cannot use.

A three-run routine

Run 1 — confirm the trigger fires. One short clip, minimal prompt, correct trigger for the checkpoint you loaded. If the motion looks no different from base H3, the trigger is missing, misspelled, or not at the front. Fix that before changing anything else.

Run 2 — load up segment two. Same shot, but write the force description properly: technique, body coordination, where the weight goes. This is the segment that separates a usable clip from a mannequin swinging its arms, and it is the one worth iterating on.

Run 3 — add the scene. Environment, camera, and on kungfu combat the style slot. Adding these last means that when the shot gets worse, you know which part did it.

If a strike still refuses to connect after run 2, switch checkpoints before you rewrite the prompt again. The two were trained differently enough that a prompt failing on one is not evidence the idea is wrong.

The bottom line

Three files, two trigger words, two templates. wushu_action at v5 is the body-mechanics model and wants technique and force described in detail. kungfu combat at 4500 steps is the staging model and wants environment, camera, and style. The higher number is a different training run, not an upgrade.

Both need the trigger at the front. After that, spend the words where they count: body mechanics on one, staging on the other.

For a hosted starting point without a local install, the MiniMax H3 generator runs the standard model in the cloud, and the H3 prompt library has reusable structures you can adapt. For prompt structure in general, see the H3 prompt format guide, and the MiniMax H3 API guide if you are driving it from code. For model background, what MiniMax H3 is.

Checkpoint names, trigger words, step counts, prompt templates, and download figures in this article come from the Jojocodex/minimax-h3-wushu-action-lora model card and repository, verified August 20, 2026. This is a community LoRA under Apache-2.0, not an official MiniMax product. License terms for the base model are summarized from the MiniMax H3 Community License.