MiniMax H3: Reference to Video
Generate a new video by feeding MiniMax H3 a reference image, up to nine images, three videos, and three audio clips alongside a text prompt, letting the model match character identity, motion, camera, voice, or style from those sources while producing native stereo audio in a single pass. Output is a single MP4 video at up to 2K resolution and 24fps for roughly 15 seconds. Ideal for character-consistent animation, style-transfer video generation, and audio-synced content creation.
Try an example
About this workflow
This workflow demonstrates MiniMax H3’s reference-to-video generation in ComfyUI: you provide a text prompt plus up to nine reference images, three short reference videos, and three audio clips, and the model produces a single MP4 with native stereo audio in one pass. The MiniMaxH3ReferenceToVideo node is the core, fusing multi-modal context (identity, motion, camera, style, and voice) with your prompt to synthesize a ~15s clip at up to 2K resolution and 24fps.
Under the hood, the Switch Model and Settings group loads the MiniMax H3 model stack via UNETLoader and CLIPLoader, while KSamplerSelect, BasicScheduler, SamplerCustomAdvanced, and BasicGuider manage sampling behavior (seeded by RandomNoise for reproducibility). ResolutionSelector sets a 16:9 size aligned to multiples of 32, and ComfyMathExpression with PrimitiveFloat handles derived values like frame count from duration and fps.
After generation, VAELoader provides both video and audio VAEs; VAEDecode reconstructs frames from video latents and VAEDecodeAudio reconstructs stereo waveforms. CreateVideo multiplexes frames and audio, and SaveVideo writes the final MP4. This makes the workflow practical for character-consistent animation, style-guided motion, and audio-synced content without separate TTS or sound design steps.
Questions & answers
Download MiniMax H3 from Comfy-Org/MiniMax-H3. Place minimax_h3_video_vae_fp16.safetensors and minimax_h3_audio_vae_fp32.safetensors in models/vae. Ensure the H3 UNet and CLIP/text encoder weights are in the correct ComfyUI model folders (e.g., models/unet, models/clip). Restart ComfyUI so UNETLoader, CLIPLoader, and VAELoader can find them.
This workflow targets about 15 seconds at up to 24fps and up to 2K resolution. Use ResolutionSelector’s 16:9 presets (multiples of 32) like 1152x640 or 1280x736. Higher resolutions, frame rates, or longer durations increase VRAM and time costs and may be unstable.
Use a small set of high-quality references (start with 1–3). Make your primary identity image the first input. Keep reference videos short and representative of the motion/camera you want. Tight, explicit prompts plus moderate guidance (CFG ~4–7) often preserve references better.
Yes. If the workflow exposes a ComfySwitchNode on the audio path, set it to bypass VAEDecodeAudio. Alternatively, leave the audio inputs empty to generate minimal audio, then replace or mute in the CreateVideo stage or in post.
Set a fixed seed in RandomNoise to reproduce the same sampling trajectory with the same inputs and settings. Changing seed, references, resolution, or sampler parameters will change the output.