🔗 H3 Relay · Audio Seam

You stitched three H3 segments together and every join does the same thing: a tiny dip, then the ambience kicks in late. It's the single most common complaint about multi-segment generation, and it isn't the model's fault. Every decoded segment starts with about 32 ms of near-silence - the decoder's priming - plus a generation transient. Trim AV then cuts the visual overlap onto that exact spot, so the join lands on nothing, then snaps to full level. About 20 ms of −12 to −13 dB hole, in a neighbourhood that was sitting at −72 dB.

H3RelayAudioSeam fixes that in the audio domain. It swaps the head of the current segment for ambient sound taken from an already-rendered segment, then crossfades back into the current segment's own audio. Length-preserving: not a single sample added or removed, so your A/V sync is untouched. That constraint is the whole design. A real crossfade can't do it - it has to eat the tail of the previous segment, and this node only touches its own input. So it patches, and the assemble layer stays a plain butt-join.

There's a second job in here that surprises people: the node also concatenates audio from segment 0 up to the current one and hands it out of the joined output - the same code path as the pack's 🧩 concat button, so no external ffmpeg pass is needed for one continuous track.

How the patch actually works

On each run it resolves the bed source from output/relay_kit/<run_id>/ - the audio sidecar every segment writes - takes a window equal in length to the patch, level-aligns it to the pre-seam audio (capped at ±6 dB), and fades back across fade_seconds. bed_select defaults to tail, so the window sits just before the seam and its level and timbre are naturally continuous with what came before. The old 0.5.0 behaviour (bed_select: quiet, take the globally quietest window) measured −42 dBFS against a −12.8 dBFS pre-seam level: it turned the seam dip into a silence hole and flattened the downbeat. It's kept only as a control.

Then an isolated-transient suppressor runs over the result - declick_ratio (4), declick_quiet_dbfs (−50), declick_max_len_ms (50). It kills one-off "tick" spikes over a dead-quiet background, and keys off how quiet the surroundings are, not how loud the spike is. That's why it doesn't shave the p/t/k plosives off your dialogue: those sit on a −35 dBFS floor, well above the gate.

The inputs worth setting

Required: audio (from Trim AV's audio output; on segment 1 there's no Trim AV, so straight from VAEDecodeAudio), run_id, stage_index.

Then two that matter:

• patch_seconds - how much of the head to replace. 0 is off and bit-exact passthrough. The tooltip suggests 2.0; the pack's own discipline table says keep it ≤ 1.2, and that's the one to follow. The patch replaces content wholesale, so dialogue sitting at the head gets eaten. The safeguard is patch_guard (on), which finds where your speech starts and shrinks the patch to 0.40 s before it - but it works off the ratio between speech energy and the surrounding median, so on material that's wall-to-wall voice or music the threshold floats above the speech peak and it never fires. Shortening the patch is the fix; the guard is backup.
• patch_guard - leave it on unless you're deliberately reproducing old behaviour.

fade_seconds (0.25) is the crossfade width at the patch boundary; bed_stage (0) picks which rendered segment supplies the bed - must be lower than stage_index; cut_head_frames should be wired to the bridge's context_frames so the suppressor judges the cropped-out head as a fresh file start. Everything under join_*, plus tile_seconds and note, is advanced and safe to ignore.

Outputs: audio (wire this to CreateVideo.audio, or to your save node), report, and joined - the running concatenated track.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ZenHG/ComfyUI-H3-Latent-Relay.git
# then restart the ComfyUI backend, which means the Python process, not the browser tab

Or ComfyUI Manager → Custom Nodes Manager → search h3-latent-relay (the registry id drops the "ComfyUI" prefix). No model downloads, no extra pip work for this node - the pack only needs torch/safetensors from your ComfyUI install and, for concatenation, av>=17, which ComfyUI already ships.

Where people get burned

Stage 1 (segment 2) needs segment 1 to have run first - the bed file is written by this node on every run, including the first segment, so run in order and don't skip stages. A "bed source missing" error means exactly that.

join_cross_ms set without join_align_seconds raises on purpose. Crossfading is overlap-add, which shortens the timeline; the node refuses rather than quietly emitting a track that drifts ahead of the picture a few ms per seam.

Last thing: this is a young pack from a solo author (ZenHG) with essentially no Reddit footprint yet, so there won't be a pile of tutorials. Read the tooltips - they're unusually honest, including where the author documents a measurement that killed his own earlier default.