⏩ H3 Relay · Finite Run

MiniMax-H3 generates 4–15 second clips. That's the model, not a setting you got wrong. So anyone making anything longer ends up in the same loop: queue segment 1, watch it, bump the stage number, rewrite the prompt to pick up the action, queue again. It works - and it's a dozen clicks per minute of footage, with you there for all of them.

⏩ H3 Relay · Finite Run turns that loop into one submit. You say how many segments you want, it builds the whole chain server-side and runs it start to finish - no frontend buttons, no per-segment edits, works from /prompt with the browser closed.

What it's really doing

The clever bit isn't the sampling. When you queue it, the node returns a graph instead of a tensor, and ComfyUI's engine registers each node in it as an ephemeral node in the same prompt. Public API, no monkey patching.

Each segment becomes a ten-node chain: MiniMaxH3ImageToVideo → H3RelayCopyBridge → RandomNoise → BasicGuider → SamplerCustomAdvanced → H3RelayLatentSave → VAEDecode / VAEDecodeAudio → H3RelayTrimAV → H3RelayJoinAV. No Latent Load either - the bridge pulls the previous segment off disk itself from run_id + stage number.

The continuation underneath is the pack's whole reason for existing. Instead of decoding segment 1 to mp4, re-encoding it and using that as an anchor - the last-frame-to-first-frame chaining everyone does with Wan - it copies the tail of the previous segment's AV latent bit-for-bit into the new one's initial latent and masks it so those frames don't get redrawn. No VAE round trip, no re-encoding drift, and audio is pinned with the picture. The pack's own seam checks read 0.000 across 7 sampling steps in hard mode.

The inputs a beginner actually sets

model, clip, vae, audio_vae and the sampling pair sampler / sigmas are shared by every segment. Then it's four things:

• prompts - one block per segment, blocks separated by a line of three or more dashes. Block 1 feeds segment 1. Fewer blocks than segments errors instead of quietly running short. One prompt for everything? Leave prompts empty and use prompt.
• segments - how many. Default 2, max 64.
• run_id - the film's name. The one people get wrong: it locates every intermediate file in output/relay_kit/<run_id>/ and every "read the previous segment". Don't reuse it for a different film.
• settle_frames - keep 0 for same-camera continuation. Set -1 for a new camera angle and it walks the hard cut back to the segment boundary. Measured: 0 gives a 0.020 frame jump; chopping 16 settle frames instead makes it 0.055, i.e. worse.

length only accepts the 5+17k ladder (5, 22, 39, 56, 73…158) and context_frames 5/22/39/56/73/90/107/124; off-ladder values raise rather than snapping.

Outputs

last_latent, images, audio, report. It isn't an output node, so nothing is saved until you save it: images and audio into CreateVideo, that into SaveVideo. last_latent is the final segment's tail if you want to continue later; report prints the plan and sidecar paths.

Install

ComfyUI Manager → Custom Nodes Manager → search h3-latent-relay (the registry id isn't the repo name; repos there can't start with "ComfyUI"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/ZenHG/ComfyUI-H3-Latent-Relay.git

Then restart the backend - refreshing the browser does nothing. No third-party node pack is needed here; only torch and safetensors, which ComfyUI ships. You do need ComfyUI ≥ 0.35.0 with MiniMax-H3 support, and that's the sharp edge: on an older build the nodes register fine and continuation just silently doesn't happen. H3 weights aren't in the pack.

Where people get burned

Memory. The joined output is one tensor resident in memory - total, not peak. At 1216×672 with 192-frame segments: 0.9 GiB for two, 4.4 GiB for ten, 16.2 GiB for thirty-seven. A 12 GB card realistically stops near ten. The node does that arithmetic from your settings and puts it in report.

No fixing it later. There's no Audio Seam or Post node in the expanded graph, and you can't eyeball segment 3 and re-roll just that one. from_segment covers recovery: set it to N and the first segment's bridge pulls stage N-1 off disk, erroring before anything loads if it's missing. It's mutually exclusive with image - your start point is somebody else's tail.

References. Use ref_image_1..3, not image. One set for the whole film is what refs_mode: lock is for; swapping per segment means the model re-recognises faces once the pinned window ends, and you get a hard cut where you wanted a seam.

Hand-writing API JSON? The frontend inserts an extra control_after_generate value right after seed, so widgets_values needs one more entry than the schema. Get it wrong and everything after seed shifts a slot - the pack's own example was rejected with HTTP 400 that way.