Minimax-H3-Video-Music-Nodes
6 Video & 2 Music nodes for the Minimax Ecosystem
Nodes (8)
Real frame anchoring, not prompt roulette
The MiniMax H3 sampler, and the fine print
Ffmpeg-backed mp4/webm export
The do-everything H3 sampler — and where its image input falls short
The 'RIFE-lite' node that's secretly just lerp — and why that's fine
Bigger H3 frames without re-rolling the seed
Music from the same checkpoint that made your video
From raw waveform to a loopable, loudness-matched track
MiniMax H3 Nodes for ComfyUI
Custom node pack wrapping the local MiniMax H3 (33.1B joint video+audio) model merged into ComfyUI core — video generation/post-processing, plus standalone MiniMax music generation/post-processing.
Requirements
- ComfyUI with the MiniMax H3 checkpoint loaded via a standard
CheckpointLoader/UNETLoader(outputs aMODELcompatible with these nodes) ffmpegon PATH (required byMiniMax H3 Video Export)- Python deps:
torch,torchaudio,numpy,Pillow(seepyproject.toml)
Installation
Drop the minimax_h3_nodes folder into ComfyUI/custom_nodes/, then restart ComfyUI. Nodes are auto-discovered — __init__.py scans every .py file in the package at load time, so no manual registration is needed.
Nodes
Video
| Node | Description |
|---|---|
| MiniMax H3 Text-to-Video Sampler | Pure text-conditioned noise-to-video sampling. Aspect-ratio presets (16:9, 9:16, 1:1, 4:3) and a motion_strength control. |
| MiniMax H3 Image-to-Video Sampler | Animates a start_image (optionally toward an end_image), encoding reference frames through a supplied VAE and splicing them into the latent before sampling. |
| MiniMax H3 Video Generate | General-purpose sampler — text conditioning with an optional image_cond input; also emits the model's joint audio track if the checkpoint exposes one. |
| MiniMax H3 Video Upscale | Spatial upscale — uses a supplied UPSCALE_MODEL if given, otherwise falls back to lanczos/bilinear/bicubic/nearest-exact. |
| MiniMax H3 Video Interpolate (RIFE-lite) | Temporal frame interpolation via linear blending, for raising effective output fps. Swap in RIFE/FILM for higher-quality results. |
| MiniMax H3 Video Export | Muxes frames (+ optional audio) to mp4/webm via ffmpeg. |
Music
| Node | Description |
|---|---|
| MiniMax Music Generate | Text/lyrics/style-audio conditioned music generation via the model's generate_music head. |
| MiniMax Music Post-Process | Trim, crossfade-loop, normalize (target LUFS), and export to wav/mp3/flac. |
Known placeholders
A few integration points are scaffolded against assumed model APIs and should be checked against your actual merged-into-core H3 implementation before relying on them in production:
model.model.decode_joint_audio(...)— used by the video samplers to pull a synchronized audio track. Falls back to silent audio if absent.model.model.generate_music(...)— required by MiniMax Music Generate; raises a clear error if the loaded checkpoint doesn't expose it.model.model.latent_channels/model.model.audio_sample_rate— read with safe defaults (16 channels, 44100 Hz) if not present on the model.
If your core build names these differently, update the relevant node file — each node is self-contained in its own .py file, so changes are isolated.
File layout
minimax_h3_nodes/
├── __init__.py # auto-discovers all node modules
├── video_text_to_video.py
├── video_image_to_video.py
├── video_generate.py
├── video_upscale.py
├── video_interpolate.py
├── video_export.py
├── music_generate.py
├── music_postprocess.py
└── pyproject.toml
License
MIT