# How to Reuse One Motion Clip for Fashion Videos in ComfyUI

Use a shared pose reference to compare short MiniMax H3 garment videos, and check product details before exporting.

By Paras Madan | 2026-11-22


Say you have photos of two jackets and want to show both with the same small turn. If you generate both clips from scratch, one jacket may get a slow turn while the other gets a moving camera. You also have to work out whether the model changed the garment or you're just looking at it from a different angle.

On November 22, developer Luan Velo [published a ComfyUI workflow](https://github.com/LuanVelo/fashion-pose-to-video/commit/6bd9a814a15553c1c3f7604a7e6e680e4c76c532) that separates the movement from the garment. It extracts one pose sequence and uses it again with each product photo. I’d try it first with two samples and one turn. You’ll learn more from that comparison than from generating a whole rack of clothes.

## Choose a turn you can check

Use two photos you own of someone wearing your sample jackets and a motion clip you have permission to use. A steady, full-body shot of one person turning slightly to the side is enough. Keep the turn short if you only have front photos. Once the person turns their back, the model has to fill in details the photo never showed.

Before opening ComfyUI, pick three details per jacket that must survive the render. For example, note the pocket count, which side the zipper pull sits on and whether the cuffs have buttons. Those details are easier to judge than asking if the generated clip looks close enough.

## Extract the motion once

The [project’s setup guide](https://github.com/LuanVelo/fashion-pose-to-video) points to three workflow JSONs. Start with `motion_W1_extract-pose_v3`. It uses SDPose to turn a section of the source clip into a moving skeleton. Give it a useful name, such as `quarter-turn`, and follow the project’s catalog step to make the pose video available as input to the next graph.

Play the extracted skeleton beside the original. Pay attention to the moment an arm crosses the torso. If the wrist jumps or the arm disappears when it should still be visible, pick a cleaner motion clip before rendering a jacket. If you reuse that broken sequence, every video in the batch starts with the same problem.

## Check the first frame before the whole clip

Open `motion_W5_quality_v11` and select the pose video and first jacket photo. The graph uses FLUX.2 klein to make the opening frame from the photo and the first pose. It then uses MiniMax H3 ref2va and pose ControlNet to animate the shot, with the jacket photo still available as a reference. The package specifies separate downloads for the models and uses native ComfyUI nodes; the developer reports testing on ComfyUI 0.46.2 with an RTX 4090.

Check the opening frame against the photo. If it already swapped the pockets or changed the cuff, fix the reference or rerun that stage before judging the animation. Repeat for the second jacket using the same pose sequence. Keep the camera framing and duration fixed so it’s easier to spot differences in the garments rather than differences in the shot.

## Inspect the turn, then upscale

Watch each clip at normal speed, then pause near the middle of the turn. Check whether the zipper moves across the chest, a pocket disappears behind the arm and comes back in a different place, or the cuff gains a button. The original, skeleton and generated video are useful side by side: a jump visible in the skeleton points to a different problem from a pocket that changes while the skeleton stays steady.

Only upscale the clips that pass. The third graph, `motion_W6_upscale_v3`, uses RelayVSR and exports a 1080 × 1920 version. Keep the earlier render for comparison. Sharper fabric can make the final video look better at phone size, but won't give you the right pocket back.

Treat this as a prototype until you’ve checked the model terms. The supplied [FLUX.2 klein base 9B checkpoint](https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9b-fp8) is gated and listed under a non-commercial license. The [MiniMax H3 files](https://huggingface.co/Comfy-Org/MiniMax-H3) have their own community agreement. For a customer-facing catalog, check that the models you use are permitted for that work before you start rendering.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams explain technical products through clear documentation, practical guides and developer-focused content.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
