MiniMax H3 Character Swap 1.5 is a ComfyUI workflow for replacing one tracked character in a source video with a new character defined by a reference image. SAM 3.1 tracks the target, the workflow cleans and maps the mask into the generation latent, and the edited subject is composited back into the source frames. The connected generation path uses MiniMax H3 Ref2VA, the h3_character_swap_v1.5 LoRA at strength 1.0, and OpenVDN DMD8 at 8 NFE.
Main features:
- SAM 3.1 video tracking, mask cleanup, dynamic subject crop, and local subject recompositing.
- MiniMax H3 video and audio latent path with the Character Swap 1.5 LoRA.
- OpenVDN/vdn-minimax-h3 stage_dmd_8nfe execution plan. This is a single-stage 8 NFE workflow; it does not include a second 4 NFE refinement or 2x upscale pass.
- H.264 MP4 output through VideoHelperSuite at the workflow's configured 24 fps.
Suggested workflow:
1. Load your source video and the replacement character reference image.
2. Set the character description and desired duration. Keep both resolution selectors consistent.
3. Change the SAM text prompt to identify only the person to replace; inspect the tracking preview before generation.
4. Confirm the H3, SAM 3.1, Character Swap 1.5, OpenVDN, MaskVidExperiments, and VideoHelperSuite dependencies are installed. Do not stack the old Turbo LoRA on the VDN8 path.
5. Run the workflow, inspect the composited result, and export the MP4.
The attached JSON has passed static graph inspection only. The VDN8 and Character Swap combination has not been independently GPU tested in this package. One auxiliary LoadImage node is bypassed; it is not part of the active transformation.
RunningHub workflow: https://www.runninghub.ai/zh-cn/post/2113278202089521154?inviteCode=rh-v1111
Bilibili tutorial: https://www.bilibili.com/video/BV1GLnH6NEPC/
Related models and resources: https://pan.quark.cn/s/9ec095b95838
AI-KSK RunningHub profile: https://www.runninghub.ai/user-center/1892566146468511746?inviteCode=rh-v1111
Description
MiniMax H3 Character Swap 1.5 is a ComfyUI workflow for replacing one tracked character in a source video with a new character defined by a reference image. SAM 3.1 tracks the target, the workflow cleans and maps the mask into the generation latent, and the edited subject is composited back into the source frames. The connected generation path uses MiniMax H3 Ref2VA, the h3_character_swap_v1.5 LoRA at strength 1.0, and OpenVDN DMD8 at 8 NFE.
Main features:
- SAM 3.1 video tracking, mask cleanup, dynamic subject crop, and local subject recompositing.
- MiniMax H3 video and audio latent path with the Character Swap 1.5 LoRA.
- OpenVDN/vdn-minimax-h3 stage_dmd_8nfe execution plan. This is a single-stage 8 NFE workflow; it does not include a second 4 NFE refinement or 2x upscale pass.
- H.264 MP4 output through VideoHelperSuite at the workflow's configured 24 fps.
Suggested workflow:
1. Load your source video and the replacement character reference image.
2. Set the character description and desired duration. Keep both resolution selectors consistent.
3. Change the SAM text prompt to identify only the person to replace; inspect the tracking preview before generation.
4. Confirm the H3, SAM 3.1, Character Swap 1.5, OpenVDN, MaskVidExperiments, and VideoHelperSuite dependencies are installed. Do not stack the old Turbo LoRA on the VDN8 path.
5. Run the workflow, inspect the composited result, and export the MP4.
The attached JSON has passed static graph inspection only. The VDN8 and Character Swap combination has not been independently GPU tested in this package. One auxiliary LoadImage node is bypassed; it is not part of the active transformation.
RunningHub workflow: https://www.runninghub.ai/zh-cn/post/2113278202089521154?inviteCode=rh-v1111
Bilibili tutorial: https://www.bilibili.com/video/BV1GLnH6NEPC/
Related models and resources: https://pan.quark.cn/s/9ec095b95838
AI-KSK RunningHub profile: https://www.runninghub.ai/user-center/1892566146468511746?inviteCode=rh-v1111