MiniMaxH3ReferenceToVideo - ComfyUI Built-in Node Documentation

MiniMax H3 Reference to Video creates the text conditioning and the empty audio-video latent needed for MiniMax H3 reference-to-video generation.

MiniMax H3 Reference to Video creates the text conditioning and the empty audio-video latent needed for MiniMax H3 reference-to-video generation. You provide a prompt plus optional reference images, videos, and audio clips, and the node encodes these references into conditioning the model can use while generating. The prompt refers to the references with <Picture i>, <Video k>, and <Audio j> tags.

Inputs

Parameter	Description	Data Type	Required	Range
clip	CLIP model used to tokenize the prompt and encode the reference media into conditioning tokens.	CLIP	Yes
vae	Video VAE used to encode reference images and reference video frames. Without it, reference images/videos only condition the text encoder.	VAE	No
audio_vae	Audio VAE used to encode reference audio. Audio is resampled to the audio VAE sample rate (32 kHz by default). Without it, reference audio only conditions the text encoder.	VAE	No
prompt	Text prompt for the video. Reference media can be addressed with <Picture i>, <Video k>, and <Audio j> tags (1-based per type). Supports multiline and dynamic prompts.	STRING	Yes
width	Width of the generated video in pixels (default: 1344).	INT	Yes	32 to 16384 (step 32)
height	Height of the generated video in pixels (default: 768).	INT	Yes	32 to 16384 (step 32)
length	Frame count at 24 fps; 124 = ~5s, trained range is ~124-362 (default: 124).	INT	Yes	5 to 3600 (step 17)
ref_image_size	Reference image sizing. match scales each reference image down only, keeping the aspect ratio, to the generation’s pixel area; max uses the reference pipeline’s 2048px short edge for best identity fidelity. Reference tokens ride through every sampling step, so max can be several times slower (default: match).	COMBO	Yes	"match" "max"
ref_images	Growable slot: connect up to 9 reference images (ref_image_1 … ref_image_9). Reference images are downscaled to a 2048px short edge if larger and are never upscaled.	IMAGE	No	0 to 9
ref_videos	Growable slot: connect up to 3 reference videos (ref_video_1 … ref_video_3). Reference video frames at 24 fps (2-15s).	IMAGE	No	0 to 3
ref_video_audios	Growable slot: connect up to 3 soundtracks (ref_video_audio_1 … ref_video_audio_3). Soundtrack of the same-numbered reference video.	AUDIO	No	0 to 3
ref_audios	Growable slot: connect up to 3 standalone reference audio clips (ref_audio_1 … ref_audio_3).	AUDIO	No	0 to 3

Notes:

• The prompt refers to reference media with 1-based tags per type: <Picture i> for images, <Video k> for videos, and <Audio j> for audio. References are presented to the model in a fixed order: images, then videos (with each soundtrack’s <Audio j> label right before its <Video k>), then standalone audio.
• A soundtrack connected to ref_video_audio_N is used with the reference video connected to ref_video_N.
• Reference videos must contain at least 5 frames (~0.2 seconds at 24 fps), otherwise the node raises an error. Frames beyond the requested length are trimmed, and the remaining frame count is adjusted to a value supported by the model.
• The requested length is aligned to a supported frame count before the latent is created.

Outputs

Output Name	Description	Data Type
positive	Conditioning containing the encoded prompt. When reference media and the relevant VAEs are provided, it also contains the encoded reference image, video, and audio content used by the MiniMax H3 model.	CONDITIONING
latent	Empty audio-video latent at the requested width, height, and length (frame count).	LATENT

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

---

Source fingerprint (SHA-256): 47df0d6d13cb02aa4f69b50a7f8d0f6c1639c1fb5e0f69bf8fc57dd4cb752db8