H3RelayInternalAcceptRaw
Accept the raw H3 shot as-is — the node that keeps your movie at native 480p
- sequence
- state
- segment
- sequence
- video
- video_path
- status
Most of the nodes in H3 Relay's internal category are things you'll never place by hand - they pop into existence when a public node expands. Accept Raw is one of the few worth understanding anyway, because it's the mechanism that makes "native 480p H3, no finishing" a first-class option instead of a hack.
The node description says it exactly: Accept native H3 frames directly, bypassing LTX enhancement and RIFE interpolation while retaining the continuation checkpoint. Feed it the current sequence, the H3 state, and the segment you just generated, and it records that segment as accepted - the checkpoint stays in the chain, so the next Generate Shot continues from it - but it skips the entire LTX/interpolation finishing path. You get the raw H3 output as your deliverable.
When you'd actually use this
Three situations. First, review loops: you want to look at the raw shot and decide whether it's worth spending LTX inference on it. Accepting it raw is the "keep it in the timeline, don't enhance yet" move. Second, a genuinely 480p deliverable - maybe it's a proof-of-concept or a mood board and you don't need the 2x finish. Third, the resumable pipeline in staged mode: the staged executor constructs per-stage prompts, and an H3-only stage can run with just the H3 model family resident. Accept Raw is what closes out that stage without pulling in the finishing models.
What it wires into
The three inputs are all internal types: sequence (H3_RELAY_SEQUENCE), state (H3_CHAIN_STATE), and segment (H3_CHAIN_SEGMENT) - the same objects the public Generate Shot thread produces. Outputs are sequence (feed to the next Generate Shot), video, video_path, and status.
There's a detail worth knowing even though you'll probably never set it: the node inherits the shot's cached output_crf from the sequence and muxes the lossless generated WAV into a raw preview MP4, stream-copying the checkpoint picture so audio lands in the preview without re-encoding the H3 video. That's the same audio trick the public nodes use; Accept Raw just does it at the accept step.
The honest take
You'll almost certainly never add this node manually. In the reference workflow it appears when the graph expands, and it's the "accept without finishing" branch that lets H3 Relay hold the sequence and enhanced streams apart. If you're reading this because you found it in an expanded graph and wondered what it was: that's it. Leave it alone, don't delete it, and let the public nodes drive. Its whole job is to make sure that skipping LTX and interpolation is a deliberate, cached, resumable choice rather than an accident.
Same install story as everything else in the pack: ComfyUI Manager search H3 Relay, or git clone https://github.com/akatz-ai/h3-relay.git into custom_nodes, then restart. Needs ComfyUI 0.32.0+, FFmpeg, and the H3 model files from MODELS.md. No extra Python dependencies, and remember the MiniMax H3 community license excludes the US, EU, UK, and Korea - check it before you start downloading the 42GB of H3 weights.
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| sequence | H3_RELAY_SEQUENCE | — | |
| state | H3_CHAIN_STATE | — | |
| segment | H3_CHAIN_SEGMENT | — |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| sequence | H3_RELAY_SEQUENCE | — |
| video | VIDEO | — |
| video_path | STRING | — |
| status | STRING | — |