H3 Relay · Assemble Raw Sequence

MiniMax H3 is the open-weight video model that generates audio and picture together in one 33B pass - genuinely good, and genuinely expensive to run. H3 Relay's whole design is built around not paying for that cost until you've decided the shot is worth it: it generates bounded shots at native 480p, writes each accepted one to disk, and only then lets you spend GPU on the LTX 2.5 enhancement and RIFE interpolation that turn raw H3 into a finished movie. Assemble Raw Sequence is the node at the end of the path where you look at those raw shots and decide you don't need any of the finishing.

It's the exit door for a chain that stays native. Wire the sequence output from your last Generate Shot into it, and it walks the chain of accepted raw shots and concatenates them into one video file - audio intact, seams clean. This is exactly what you want for FastH3 profile chains (which skip the finishing stream by design) or for any standard H3 run where you're happy at 480p with the generated audio and would rather save the VRAM than upscale.

How it works

The node takes a single sequence input of type H3_RELAY_SEQUENCE - the same wire every Generate Shot hands down the chain. Nothing is regenerated here. H3 Relay stages are disk-backed, content-addressed cache records, so this node reads the accepted shot artifacts and assembles them. Continuation overlap between shots was already trimmed by Generate Shot, so you don't get doubled frames at the joints.

Audio is where the plumbing shows. H3 keeps the checkpoint picture and the lossless generated WAV as separate internals. This node stream-copies the picture and muxes the WAV in, which means the VIDEO output carries sound without re-encoding H3's frames - but it also means FFmpeg needs to be on your PATH, or the mux dies. The video lands under ComfyUI's protected user/__h3_relay_cache area for new run names; connect a native Save Video node to any stage's VIDEO output if you want it published under output/.

The inputs that matter

Three inputs, and honestly only one needs touching:

• sequence (required) - the H3_RELAY_SEQUENCE from your last Generate Shot.
• filename (default h3_relay_raw_%date:yyyy-MM-dd_HH-mm-ss%) - the output name; the %date% placeholder keeps runs from clobbering each other.
• audio_bitrate (INT, 64–512, default 256 kbps) - how much room the muxed audio gets. 256 is fine for generated speech and SFX; go higher if you're keeping the file as a master.

It's an output node, so it shows up as a job in the queue. Outputs are video (VIDEO, for preview or Save Video), video_path (STRING - where the file actually landed), and status (STRING - success state).

Install

Search H3 Relay in ComfyUI Manager and restart, or manually:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/akatz-ai/h3-relay.git

You need ComfyUI 0.32.0+, FFmpeg on PATH, and MiniMax H3's model files - FL2VA and Ref2VA checkpoints, the Qwen3-VL 32B text encoder, and the video + audio VAEs - from Comfy-Org/MiniMax-H3. The README's MODELS.md has the exact hf download commands. One wrinkle worth knowing before you start: the MiniMax H3 Community License geofences the weights out of the US, EU, UK and South Korea. The pack doesn't ship any weights and doesn't grant you that license - that part is between you and MiniMax.

Where people get burned

• You have interpolation nodes in the graph. This node assembles the raw sequence stream, not the enhanced one. If you enabled Interpolate nodes and wire into Assemble Raw, you're leaving their work on the table - bypass the interpolation nodes to assemble LTX video at 24fps, or use the main Assemble node on the enhanced stream instead.
• Missing FFmpeg. Assembly fails at the mux step, not at generation. ffmpeg -version before you blame the node.
• Expecting it to override your runs. It reuses accepted artifacts; if you want different footage, reroll at Generate Shot, not here. Changing a shot invalidates only the finishing derived from it, so a reshot chain reassembles cheaply.

The name is a bit of a lie in one direction - it does publish the file, not just preview it. Treat it as your raw-release node, and pair it with the FastH3 VSA loader when you want fast, rough-and-ready H3 chains without the finishing pipeline at all.