VRGDG MiniMax H3 Reference Media From Paths
Point it at your reference files and let it fill the H3 sockets
- ref_image_0
- ref_image_1
- ref_image_2
- ref_image_3
- ref_image_4
- ref_image_5
- ref_image_6
- ref_image_7
- ref_image_8
- ref_video_0
- ref_video_1
- ref_video_2
- ref_video_audio_0
- ref_video_audio_1
- ref_video_audio_2
MiniMax H3 is a reference-driven model: you can hand it up to nine reference images and three reference videos, and it conditions the generation on them. That's powerful, and also a wiring nightmare in a big builder workflow - nine image sockets plus three video sockets plus three audio sockets is a lot of dangling cables. VRGDG MiniMax H3 Reference Media From Paths is the pack's "builder-friendly" answer: type (or paste) the file paths, and it loads and sorts the media into the right H3 reference slots for you, returning None for any slot you didn't use.
How it works
Two text inputs, both multiline. image_paths accepts up to nine reference images - as a JSON list, as an object with image_paths/images keys, or simply one path per line. video_references accepts up to three reference videos, either as a JSON list of paths or as objects that can carry path, start_seconds, duration, and use_audio - which is how you grab a specific segment of a longer video instead of always using its head.
The outputs are the payoff. Nine ref_image_0 … ref_image_8 sockets, three ref_video_0 … ref_video_2 sockets, and three ref_video_audio_0 … ref_video_audio_2 sockets - one for each reference video's audio track, which H3 can use for audio conditioning and lipsync. The crucial detail is the ordering: the tooltip is explicit that these are ordered references, so position in your list maps directly to which H3 reference slot it lands in. If image #2 in your list is the character's face, that face is ref_image_1 - get the order wrong and the model will mix up who's who.
Unused slots return None, which is the builder-friendly bit: you can wire the full set of outputs once and leave the empties dangling without the graph erroring out.
The inputs that matter
image_paths- up to 9, ordered; one per line or JSON.video_references- up to 3, ordered; addstart_seconds/durationto trim,use_audioto pull the track.
Install
Pack-wide: ComfyUI Manager → search vrgamedev (or add https://github.com/vrgamegirl19/comfyui-vrgamedevgirl), restart, hard-refresh. Manual: git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl into custom_nodes, pip install -r requirements.txt; Windows portable needs Cython + scikit-build-core installed first.
Troubleshooting
Path problems are the top of the list. These are filesystem paths - relative ones resolve against ComfyUI's input directory, and a typo returns an empty slot rather than an error, so a missing reference can silently change a generation. JSON syntax errors are the other classic; if you're hand-writing the video list, use one path per line instead until the start_seconds/duration trimming is actually needed. And if your references come back out of order, it's your list's order, not the node - the mapping is positional by design.
Inputs (2)
| Name | Type | Default | Description |
|---|---|---|---|
| image_paths | STRING | [] | Ordered MiniMax H3 reference images. Use a JSON list, an object with image_paths/images, or one path per line. Supports up to 9 images. |
| video_references | STRING | [] | Ordered MiniMax H3 reference videos. Use a JSON list of paths or objects with path, start_seconds, duration, and use_audio. Supports up to 3 videos. |
Outputs (15)
| Name | Type | Description |
|---|---|---|
| ref_image_0 | IMAGE | — |
| ref_image_1 | IMAGE | — |
| ref_image_2 | IMAGE | — |
| ref_image_3 | IMAGE | — |
| ref_image_4 | IMAGE | — |
| ref_image_5 | IMAGE | — |
| ref_image_6 | IMAGE | — |
| ref_image_7 | IMAGE | — |
| ref_image_8 | IMAGE | — |
| ref_video_0 | IMAGE | — |
| ref_video_1 | IMAGE | — |
| ref_video_2 | IMAGE | — |
| ref_video_audio_0 | AUDIO | — |
| ref_video_audio_1 | AUDIO | — |
| ref_video_audio_2 | AUDIO | — |