MiniMax H3 Video Export

Every sampler in this pack hands you a latent and audio, but a latent isn't something you can post anywhere. MiniMaxH3VideoExport is the last node in the graph - the one that takes decoded frames (plus optional audio) and muxes them into an actual mp4 or webm file on disk. It's an output node, the "save video" button, and it's where your H3 clip finally becomes a file.

The mechanism is refreshingly boring, which is a compliment. The node writes your frames to a temp folder as PNGs, then shells out to ffmpeg with -framerate set to your chosen fps, and encodes to libx264 (mp4) or libvpx-vp9 (webm) at your crf, forced to yuv420p so the file plays everywhere. If you wire an AUDIO in - like the joint_audio socket coming out of any H3 sampler - it adds that as a second input, encodes it as AAC, and uses -shortest so the audio doesn't overrun the video. The output lands in ComfyUI's normal output directory with an auto-incrementing counter, and the node reports it back to the UI for preview.

Inputs that matter:

• images (required) - wire this from a VAE Decode on your sampler's video_latent.
• fps (1–120, default 24) - set it to the frame rate you actually generated at. If you used the Interpolate node to double your frames, this is where you declare the new rate.
• format - mp4 (H.264, plays everywhere) or webm (VP9, smaller, but pickier about where it plays).
• crf (0–51, default 19) - lower is higher quality and bigger file. 19–23 is the sane range for H.264; the widget caps at 51 which is also fine for VP9.
• filename_prefix - the base name before the counter.
• audio (optional) - the sampler's joint_audio.

Install. This is the one node in the pack with a real system dependency: ffmpeg must be on your PATH. On Windows that usually means installing it and adding it to PATH yourself; on Linux/macOS it's apt install ffmpeg / brew install ffmpeg. The pack's Python deps are just torch, torchaudio, numpy, Pillow:

cd ComfyUI/custom_nodes
git clone https://github.com/TensorVizion/Minimax-H3-Video-Music-Nodes

or install through ComfyUI Manager, then restart.

Where it bites. The big one: if ffmpeg isn't reachable, the node fails with a subprocess error at export time, not at install time - so test it early with a two-frame clip before you spend an hour generating. Second, remember the pack's audio caveat: if your H3 checkpoint doesn't expose decode_joint_audio, the sampler hands you a silent audio track, and this node will happily mux that silence into your mp4. An "exported with sound" video can quietly have no sound at all - check with generate_audio off if you don't actually need the track. Third, webm + audio + VP9 is slower to encode than mp4, so for quick iteration mp4 at crf 23 is the pragmatic default. And as with everything H3, the ~42.5GB model with no consumer-VRAM floor is the real bottleneck in the room - this node is the cheap, reliable part.