BUNNY H3 Ref2VA Motion Transfer
This ComfyUI workflow is used to replace the characters in a reference video with a reference image character, while trying to preserve the original video's motion, camera movement, cuts, and rhythm. It supports automatically converting the reference video to 24 fps, taking 720P or 480P as input resolution, automatically recognizing landscape and portrait orientations, and outputting through primary and secondary sampling.
Instructions for Use
1. Import the JSON into ComfyUI and first handle missing nodes and model files.
2. Upload the original video in the BUNNY|Reference Video node.
3. Place the target character images sequentially into the reference image group. Image 1 is enabled by default; when using Image 2 or more, manually enable the corresponding groups.
4. Select the reference video resolution in BUNNY|720P On / 480P Off: enabled means a short edge of 720 pixels, disabled means a short edge of 480 pixels. Landscape and portrait orientations are automatically determined by the original video.
5. Write prompts for the current clip and paste them into the BUNNY|Six-Segment Prompt node. The attached Skill can assist in checking the complete video and generating six segments of content: subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, non_diegetic_music.
6. Check whether <Picture N> in the prompts matches the image input order, then run, and check the primary and secondary sampling results respectively.
The Skill package will be released later in a Bilibili video.
Required Nodes
- ComfyUI: Native H3 reference generation, calculation, scaling, sampling, and resolution nodes.
- ComfyUI-VideoHelperSuite: Video Load, Load Video Info, Save MP4.
- ComfyUI-KJNodes: Set/Get nodes and H3 SageAttention patch.
- ComfyUI-Easy-Use: easy float node.
- MiniMax H3 Latent Upscaler: MinimaxH3LatentUpscaler3D node used for secondary sampling.
- SageAttention: Required for the currently enabled SageAttention patch; the version must be compatible with your graphics card and ComfyUI environment.
The workflow also retains the "ignore multi-group isolated sea" group control node from the original environment. I could not find a verifiable public installation address. If missing upon import, you can remove these auxiliary nodes and manually manage the enabled status of each reference group.
Notes
- **720P/480P is the reference video resolution sent into H3, not the primary sampling output resolution.** The generation size is controlled separately by the output pixel budget and secondary sampling settings.
- The workflow converts the input video to 24 fps and automatically rounds down the duration to the frame count of 17n + 5 supported by H3, so the end of the video may be short by up to 16 frames. Check the actual generated duration for important endings.
- The original video's audio track is not connected to H3's video/audio reference input. Music or original sound needs to be composited in post-production.
- When two people cross quickly or occlude each other, prompts should bind the target character according to the original character's continuous motion trajectory, and clearly write the corresponding relationship after swapping. If identity mixing still occurs when inputs are correct, you can test with a different seed first.
- Models, character images, and the original video must be provided by yourself. This JSON has undergone structure and connection checks; I have not independently verified the complete output in another freshly installed ComfyUI environment.
Source Attribution
Keep videos under 10 seconds. Dense cut footage (such as 5 cuts in 6 seconds) needs to be split into two segments. Complex actions like those in the demo video require using motion repair LoRA to ensure that rotations and rolls do not break. Skill usage method: install the Skill to your AI, give it the video and reference images, and it will give you the prompts. If character mixing or failure issues occur in the video, first interrogate your AI! Make it open its eyes and carefully examine the video and anchor character identity words
The motion transfer basic node graph is adapted from the original workflow video shared by 吃猪侠233.
BUNNY added prompt methods, input instructions, reference video automatic preprocessing, layout, and output naming.
Video Motion Transfer_Auto Landscape and Portrait_720p480p_24fps
BUNNY H3 Ref2VA Motion Transfer
This ComfyUI workflow is used to replace the characters in a reference video with a reference image character, while trying to preserve the original video's motion, camera movement, cuts, and rhythm. It supports automatically converting the reference video to 24 fps, taking 720P or 480P as input resolution, automatically recognizing landscape and portrait orientations, and outputting through primary and secondary sampling.
Instructions for Use
1. Import the JSON into ComfyUI and first handle missing nodes and model files.
2. Upload the original video in the BUNNY|Reference Video node.
3. Place the target character images sequentially into the reference image group. Image 1 is enabled by default; when using Image 2 or more, manually enable the corresponding groups.
4. Select the reference video resolution in BUNNY|720P On / 480P Off: enabled means a short edge of 720 pixels, disabled means a short edge of 480 pixels. Landscape and portrait orientations are automatically determined by the original video.
5. Write prompts for the current clip and paste them into the BUNNY|Six-Segment Prompt node. The attached Skill can assist in checking the complete video and generating six segments of content: subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, non_diegetic_music.
6. Check whether <Picture N> in the prompts matches the image input order, then run, and check the primary and secondary sampling results respectively.
The Skill package will be released later in a Bilibili video.
Required Nodes
- ComfyUI: Native H3 reference generation, calculation, scaling, sampling, and resolution nodes.
- ComfyUI-VideoHelperSuite: Video Load, Load Video Info, Save MP4.
- ComfyUI-KJNodes: Set/Get nodes and H3 SageAttention patch.
- ComfyUI-Easy-Use: easy float node.
- MiniMax H3 Latent Upscaler: MinimaxH3LatentUpscaler3D node used for secondary sampling.
- SageAttention: Required for the currently enabled SageAttention patch; the version must be compatible with your graphics card and ComfyUI environment.
The workflow also retains the "ignore multi-group isolated sea" group control node from the original environment. I could not find a verifiable public installation address. If missing upon import, you can remove these auxiliary nodes and manually manage the enabled status of each reference group.
Notes
- **720P/480P is the reference video resolution sent into H3, not the primary sampling output resolution.** The generation size is controlled separately by the output pixel budget and secondary sampling settings.
- The workflow converts the input video to 24 fps and automatically rounds down the duration to the frame count of 17n + 5 supported by H3, so the end of the video may be short by up to 16 frames. Check the actual generated duration for important endings.
- The original video's audio track is not connected to H3's video/audio reference input. Music or original sound needs to be composited in post-production.
- When two people cross quickly or occlude each other, prompts should bind the target character according to the original character's continuous motion trajectory, and clearly write the corresponding relationship after swapping. If identity mixing still occurs when inputs are correct, you can test with a different seed first.
- Models, character images, and the original video must be provided by yourself. This JSON has undergone structure and connection checks; I have not independently verified the complete output in another freshly installed ComfyUI environment.
Source Attribution
Keep videos under 10 seconds. Dense cut footage (such as 5 cuts in 6 seconds) needs to be split into two segments. Complex actions like those in the demo video require using motion repair LoRA to ensure that rotations and rolls do not break. Skill usage method: install the Skill to your AI, give it the video and reference images, and it will give you the prompts. If character mixing or failure issues occur in the video, first interrogate your AI! Make it open its eyes and carefully examine the video and anchor character identity words
The motion transfer basic node graph is adapted from the original workflow video shared by 吃猪侠233.
BUNNY added prompt methods, input instructions, reference video automatic preprocessing, layout, and output naming.
BUNNY H3 Ref2VA Motion Transfer
This ComfyUI workflow is used to replace the characters in a reference video with a reference image character, while trying to preserve the original video's motion, camera movement, cuts, and rhythm. It supports automatically converting the reference video to 24 fps, taking 720P or 480P as input resolution, automatically recognizing landscape and portrait orientations, and outputting through primary and secondary sampling.
Instructions for Use
1. Import the JSON into ComfyUI and first handle missing nodes and model files.
2. Upload the original video in the BUNNY|Reference Video node.
3. Place the target character images sequentially into the reference image group. Image 1 is enabled by default; when using Image 2 or more, manually enable the corresponding groups.
4. Select the reference video resolution in BUNNY|720P On / 480P Off: enabled means a short edge of 720 pixels, disabled means a short edge of 480 pixels. Landscape and portrait orientations are automatically determined by the original video.
5. Write prompts for the current clip and paste them into the BUNNY|Six-Segment Prompt node. The attached Skill can assist in checking the complete video and generating six segments of content: subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, non_diegetic_music.
6. Check whether <Picture N> in the prompts matches the image input order, then run, and check the primary and secondary sampling results respectively.
The Skill package will be released later in a Bilibili video.
Required Nodes
- ComfyUI: Native H3 reference generation, calculation, scaling, sampling, and resolution nodes.
- ComfyUI-VideoHelperSuite: Video Load, Load Video Info, Save MP4.
- ComfyUI-KJNodes: Set/Get nodes and H3 SageAttention patch.
- ComfyUI-Easy-Use: easy float node.
- MiniMax H3 Latent Upscaler: MinimaxH3LatentUpscaler3D node used for secondary sampling.
- SageAttention: Required for the currently enabled SageAttention patch; the version must be compatible with your graphics card and ComfyUI environment.
The workflow also retains the "ignore multi-group isolated sea" group control node from the original environment. I could not find a verifiable public installation address. If missing upon import, you can remove these auxiliary nodes and manually manage the enabled status of each reference group.
Notes
- **720P/480P is the reference video resolution sent into H3, not the primary sampling output resolution.** The generation size is controlled separately by the output pixel budget and secondary sampling settings.
- The workflow converts the input video to 24 fps and automatically rounds down the duration to the frame count of 17n + 5 supported by H3, so the end of the video may be short by up to 16 frames. Check the actual generated duration for important endings.
- The original video's audio track is not connected to H3's video/audio reference input. Music or original sound needs to be composited in post-production.
- When two people cross quickly or occlude each other, prompts should bind the target character according to the original character's continuous motion trajectory, and clearly write the corresponding relationship after swapping. If identity mixing still occurs when inputs are correct, you can test with a different seed first.
- Models, character images, and the original video must be provided by yourself. This JSON has undergone structure and connection checks; I have not independently verified the complete output in another freshly installed ComfyUI environment.
Source Attribution
Keep videos under 10 seconds. Dense cut footage (such as 5 cuts in 6 seconds) needs to be split into two segments. Complex actions like those in the demo video require using motion repair LoRA to ensure that rotations and rolls do not break. Skill usage method: install the Skill to your AI, give it the video and reference images, and it will give you the prompts. If character mixing or failure issues occur in the video, first interrogate your AI! Make it open its eyes and carefully examine the video and anchor character identity words
The motion transfer basic node graph is adapted from the original workflow video shared by 吃猪侠233.
BUNNY added prompt methods, input instructions, reference video automatic preprocessing, layout, and output naming.

