Fight Flow V1 is an experimental MiniMax H3 LoRA aimed at making an action exchange continue from one move into the next. Its author focuses on how characters respond to each other: an attack changes distance or balance, the opponent reacts, and that result becomes the starting state of the following action.
The published model card includes a downloadable H3_Fight_Flow_V1.safetensors adapter and a demonstration. We reviewed that card on October 10, 2026. The improvements below are author claims, not results measured by H3Vid. We have not run the adapter, and the hosted MiniMax H3 LoRA generator currently uses a different preset.
What Fight Flow tries to change
A common action-video problem is that characters perform isolated moves: one attack finishes, everyone returns to a neutral stance, and the next attack starts as if the previous one never happened. Fight Flow's stated goal is to preserve the consequences of contact, deflection and repositioning.
That is a different target from teaching a specific technique or accelerating inference. The author describes a sequence of initiation, response, contact, consequence and a change of initiative. In a multi-character scene, another opponent may join after the current exchange rather than everyone moving independently at once.
This gives you a concrete viewing criterion. After a block, do the characters remain in the positions created by that block? Does the next movement make sense from those positions? A dramatic camera cut can disguise a reset, so an uninterrupted shot is more informative for an initial continuity check than a montage.
Select the trigger by the interaction
The model card gives three triggers. Choose the one that describes the dominant relationship between the performers, rather than simply naming an object that happens to appear in the scene.
| Trigger | Documented focus |
|---|---|
Unarmed combat interaction. | Punches, kicks, grappling, throws and body control |
Melee weapon interaction. | Close-combat weapon lines, blocks and contact |
Gun-fu interaction. | Firearm handling and close-range firearm interaction |
For example, a scene with a staff resting against a wall may still be an unarmed exchange. Conversely, a choreographed staff duel depends on how the weapon lines meet, not just whether two people hold staffs. The trigger describes the interaction category; it is not a guarantee that hands, weapons and contact points stay correct in every frame.
Keep the first scene easy to read. Two adult performers, contrasting rehearsal clothing, a plain studio and a fixed wide camera make it easier to distinguish coordinated action from independent movement. Complex backgrounds, many opponents and rapid camera motion introduce extra variables before you know whether the adapter addresses the problem you care about.
Understand the author's two-pass weights
The card reports tested weights of 0.65 for the first pass and 0.25 for the second pass. Those numbers describe the author's recipe. The card does not provide a complete downloadable ComfyUI graph, sampler configuration or universal prescription for every H3 checkpoint.
A second-pass weight is meaningful only when your workflow actually has a second sampling stage. Do not turn these values into two simultaneous adapters or assume a one-pass graph should use both. Preserve your working model and LoRA loaders, and verify the adapter's compatibility with your chosen base before changing the sampler stack.
The author recommends treating Fight Flow as an interaction layer alongside specialized action and motion-repair LoRAs. That describes an intended combination, not evidence that arbitrary adapter stacks are compatible. Each added adapter changes the experiment. Begin with a documented, working local recipe and introduce one change at a time; keep its weights and revision beside the saved result.
A prompt that makes continuity inspectable
Here is an original editorial prompt for a controlled rehearsal scene. It is not the author's showcase prompt, and we have not generated or tested it:
Unarmed combat interaction. Two adult stunt performers rehearse a short non-injurious exchange in a bright training studio. The performer in blue initiates a controlled movement; the performer in grey responds and redirects it. Both keep the distance and posture created by that response as they reposition for the next movement. One continuous wide shot, fixed camera, both performers fully visible, no cuts, no blood.The important editorial choice is to make the previous action's consequence visible. Start with a small exchange rather than asking for a catalogue of moves. Once the scene is readable, change one requirement, such as camera movement or an additional performer, and inspect whether continuity survives.
Review interaction and image quality separately
Compare the same scene with and without the adapter while keeping the checkpoint, seed, duration and other settings fixed where your workflow permits. Watch the clip at normal speed, then inspect moments immediately before and after contact.
Look for the relationship between bodies: do they react to the same event, or does one respond before the other acts? Check whether their relative distance suddenly changes. Inspect hands and limbs for intersections, and follow each performer through the entire exchange. If a weapon is part of the choreography, examine its shape and ownership as a separate visual issue.
Record continuity, anatomy, identity and overall image quality separately. A more connected exchange may still introduce a distracting visual defect. Keep unsuccessful attempts in the comparison set; the card's selected demonstration cannot establish a success rate for your prompts.
Training hardware does not establish inference requirements
The author reports training on an RTX 6000D, around 60GB peak VRAM and a longer training process than their usual LoRA work. Those are training observations. They do not specify the memory needed to generate your chosen clip with an already trained adapter.
The card also describes D-OPSD training in which a teacher can reference target video. Treat that as the author's explanation of the training method, not an independently reproduced reason for better results. Your practical next step is to inspect the current files and documentation, then evaluate one controlled scene in your existing ComfyUI setup. The published card alone does not establish hosted-provider compatibility or a ready-made online action tool.
The cover is an AI-generated concept illustration, not a Fight Flow demonstration or MiniMax H3 output.