BUNNY H3 Ref2VA Motion Transfer

Replace characters in a reference video while following its motion, camera movement, cuts, and timing. This ComfyUI workflow includes automatic 24 FPS video input, a selectable 720p/480p reference resolution, source aspect-ratio detection, and a two-pass generation pipeline.

The download includes the workflow JSON and my H3 Compact Motion Transfer prompt skill. Model weights and sample media are not included.

🚀 Run It on RunningHub

🎁 New users on RunningHub International can get 1000 RH Coins after registration.

https://www.runninghub.ai/zh-cn/post/2106347185452457985/?inviteCode=rh-v1679

国内用户可用：https://www.runninghub.cn/post/2106556243405004802/?inviteCode=cn-v1093

How to use

1. Import the workflow JSON into ComfyUI and resolve any missing nodes or model selections.

2. Upload your source clip to BUNNY | Source video (24 fps).

3. Upload each target character image to its corresponding reference-image group. Picture 1 is enabled by default; enable Picture 2 and any additional image groups you use.

4. Set BUNNY | 720P On / 480P Off. On uses a 720-pixel short edge for the reference video; off uses 480 pixels. Portrait or landscape orientation is detected from the source.

5. Write a prompt for the actual clip and paste it into BUNNY | Six-section prompt. The attached skill can inspect the clip and produce the six sections: subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, and non_diegetic_music.

6. Check that every <Picture N> in the prompt matches the image input order. Run the workflow and review the first-pass and second-pass outputs.

The skill ZIP is a prompt-writing aid. It does not install ComfyUI nodes or automatically feed text into the workflow.

Required nodes.

• ComfyUI-VideoHelperSuite — video loading, video information, and MP4 output.

• ComfyUI-KJNodes — Set/Get nodes and the H3 SageAttention patch.

• ComfyUI-Easy-Use — easy float.

• MiniMax H3 Latent Upscaler — the second-pass MinimaxH3LatentUpscaler3D node.

• SageAttention — required by the active SageAttention patch; install a version compatible with your GPU and ComfyUI environment.

The saved layout also contains an original-platform group-control helper named 忽略多组孤海. I could not verify a public installation link for it. It has no generation-data connections; if it is missing, remove those helper nodes and manage the reference groups’ enabled states manually.

Model files used by the saved JSON

• Base model: minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors — hybrid model repository.

• Text encoder: qwen3vl_32b_minimax_h3_ultra_uncensored_heretic_int8_convrot.safetensors — encoder repository.

• Video VAE: minimax_h3_video_vae_int8_convrot.safetensors — Comfy-Org MiniMax H3 models.

• Turbo LoRA: minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors — LightX2V Turbo models.

• Latent upscaler: the saved selector names minimax_h3_latent_upscaler_3d_fp16_V2.safetensors. Check the upscaler weights repository and select the compatible checkpoint you actually installed.

• Motion LoRA: the saved selector names Motion_RepairV2.safetensors.

Notes

• 720p/480p describes the video fed into H3 as a reference. It is not the first-pass output resolution. Output size is controlled separately by the pixel-budget and second-pass settings.

• The workflow converts the input to 24 FPS. Its automatic duration calculation rounds down to an H3-supported 17n + 5 frame count, so it may omit up to 16 frames from the end. Check the resulting duration when the final moment matters.

• The source video's audio is not connected as an H3 audio reference. Add music or the original soundtrack during editing.

• Fast crossings and occlusion can confuse character identity. Map each target to a continuous source performer, and state who follows whom after a position swap. If a run drifts despite correct inputs, try another seed before lengthening the prompt.

• Model weights, character images, and source videos must be supplied separately. This JSON has been checked for structure and wiring; a clean-install render on another ComfyUI setup has not been independently verified.

Credits

The underlying H3 first-pass and second-pass graph is adapted from the workflow shared by 吃猪侠233: original workflow video. BUNNY added the prompt workflow, input guidance, automatic reference-video preprocessing, layout, and output naming.

My practical tips

• Keep each reference clip under 10 seconds.

• Split footage with dense cuts into smaller segments. For example, a six-second clip containing five shots is better handled as two separate generations.

• For complex movement such as the rotations and rolls in my demo, I use the Motion Repair LoRA to help maintain the motion.

• To use the attached skill, install it in your AI assistant, then give it the source video and character reference images. It will inspect the clip and write a prompt for that specific segment.

• If the characters swap identities or the replacement fails, put your AI in the interrogation room first. Make it watch the entire clip again and prove it knows who is who across every cut, crossing, and camera move.