🔗 H3 Relay · Copy Bridge

By default, if you want segment 2 to pick up where segment 1 ended, you do pixel continuation: decode the mp4, encode the last frame back through the VAE, and feed that as an anchor keyframe. Three costs: the round-trip is lossy; the anchor and the sampling latent are different kinds of object, so the continuation drifts; and you can only anchor frame 0.

H3RelayCopyBridge takes the other road. MiniMax H3 generates picture and audio as one packed AV latent, and ComfyUI's H3 support exposes a native minimax_keyframes channel that accepts anchors at any resolved frame index. So instead of re-encoding anything, the bridge copies the previous segment's latent tail bit for bit into the head of the current segment's initial latent, and writes a noise mask over that region. Pinned means zero redraw: those rows get their velocity scaled by the mask and never get re-sampled. Nothing is decoded - the only disk artifact is a ~12 MB latent sidecar per segment.

This is the pack's one bridge. Since 0.6.0 the old conditioning-only route was folded in as a second, parallel path - the two act on different objects (latent vs conditioning), so they don't conflict:

| | [0] latent copy (motion) | [3] conditioning pin (framing) | |---|---|---| | mechanism | bit-exact tail copy, pinned rows zero redraw | tail pinned into conditioning, model redraws | | blur on reproduce | none | yes | | brightness step at seam | large (~0.040) | small (~0.004) |

Connect both and you get the copy route's zero-blur with the pin route's low brightness step: measured at 0.3 MP, the seam brightness step is 0.0009 with the composite bridge vs 0.0097 with conditioning alone, about 11× better.

Wiring it

Required: latent (this segment's initial AV latent, from the official loader or the pack's own), context_latent (the previous segment's full latent, from Latent Load), context_frames.

Connect context_latent on segment 1 too. Latent Load hands over an empty context when stage_index is 0 and the bridge passes through automatically - no continuation, no trimming. Bypassing the wire because "there's no previous segment" removes a required input, and the whole prompt gets rejected with Required input is missing.

context_frames default 22. The H3 VAE's temporal kernel is (1,4,4,4,4) - five latent tokens span seventeen pixel frames - so only 5 + 17k windows are legal: 5, 22, 39, 56, 73, 90, 107, 124. Off-grid values raise rather than snapping, because a misaligned window renders a spatially shifted seam. It must also be smaller than the segment length.

Outputs: latent → your sampler's latent input (it must go through the bridge), trim_frames (INT) → Trim AV's trim_frames, conditioning → the sampler's positive, report.

mask_mode stays hard for real work. ramp, blend, window and taper exist, and the author's note is blunt: a retest with better metrics found all four modes still jump at the seam, they just fail differently - hard keeps framing but spikes the motion, ramp/blend smooth the motion and rewrite the framing. Tuning the mask alone doesn't fix a seam; that's why the composite bridge exists.

Optional extras that earn their slot: pin_audio (on - carries the previous segment's audio tail into this segment's audio latent as sampling context), ref_anchor_stage/ref_anchor_latent (a global appearance anchor riding every step at near-zero noise, an attention sink against long-range drift), and voice_anchor - feed a LoadAudio → VAEEncodeAudio latent of this segment's speaker, ≥0.9 s, and timbre stops cross-contaminating across a speaker change. An empty or silent anchor raises immediately; to disable it, unplug the wire, don't hand it a zero latent.

Install

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

Or ComfyUI Manager → search h3-latent-relay. Restart the backend. The hard requirement is the host, not a pip package: ComfyUI ≥ 0.35.0 with the MiniMax-H3 extras. On an older host the nodes still register and the seam is silently not continued, which is the worst possible outcome - you only catch it by looking. The pack prints an explicit banner line if H3 support is missing.

The two failures you'll actually hit

A shape mismatch deep inside the model means the anchor grid doesn't match the target grid. minimax_keyframes is computed against the target spatial grid - so if you ran a latent upscale before the bridge and then connected conditioning, you get something like shape mismatch [2392,96] vs [3134,96]. Only the native-resolution pass connects the bridge's conditioning; the second pass's guider takes the prompt encoder's positive directly, and the latent-side pin (prefix copy + noise mask) is grid-agnostic and still works.

"stage_index=N says this is segment N+1 but there's no previous segment to continue" - the run_id and stage_index on the save and the load don't agree, or you changed films without renaming. Both are hard errors, deliberately. One caveat: this pack is new, from a single author, and there's no community folklore for it yet - the H3 ecosystem itself only dates from August 2026. The tooltips are long and mostly measurements, which is more than most packs give you. Read them.