🔗 H3 Relay · Post

Most ways of hiding a bad seam are lies. Cross-dissolve two frames and you've split one jump into two visible jumps. Crop the blurry settle frames and you've traded a soft blur for a visible skip. Both of those change the timeline, which means they can't be undone and they compound across segments.

H3RelayPost is the honest alternative: it works purely in the image domain, on already-decoded frames. Frame count unchanged, audio untouched, no extra sampling. If you leave every dial at 0 - which is the default - it is a bit-exact pass-through and costs you nothing but a graph node. When it does fire, it's applying statistics and convolution, not generating anything, so whatever it does is deterministic and instant. That's the argument the post-processing essay makes for the whole layer: reach for the cheap deterministic primitive before the expensive generative one.

It goes after Trim AV, taking the cropped images, and it's one of the optional four in the pack - delete it and the film still renders.

The knobs, grouped by what they're actually for

Required: just images - wire it from Trim AV's images.

Cross-segment (needs guide): match_prev is a Reinhard-style statistics match - first and second moments, per channel - aligning the head of this segment to the seam's other side. lowfreq_pull is the gentler additive version, aligning only the low-frequency grade. Both are off. Both are also gated by cross_seg_ack, which defaults false and makes them abstain entirely. That gate is the most interesting thing in this node, and I'll come back to it.

Within-segment (no guide needed): hist_match aligns the head's tonal distribution to the segment body, wb_match fixes the R:G:B ratio - these two are orthogonal, one handles level, one handles colour temperature. Both act over head_zone_frames (24) with a linear falloff, and both are off.

High frequency: deconv_strength is Wiener deconvolution - inverse-filtering under an assumed PSF, which can genuinely recover detail rather than just amplifying it, at the risk of ringing if you push it. detail_borrow swaps the head's high-frequency structure for the body's. Complementary: one restores, one transplants. Both off.

The blur-zone fix: settle_auto (and the cruder settle_sharpen) is the one you'll actually want. Redrawn heads come back soft for the first ~15 frames. These apply an unsharp pass over settle_sharpen_frames (24) and, with settle_auto, measure each frame's sharpness deficit against the segment body and scale the compensation to it - the measured curve isn't linear, which is why the fixed settle_sharpen never matched it well. Try settle_auto at 0.5–1.0. Honest limitation, stated by the author: sharpening recovers contrast, not detail that's gone. If the structure is there and merely soft, it helps.

Outputs: images and report. Wire the guide from Trim AV's prev_tail - the last frame of the pinned region, i.e. the other side of the seam. Without it, the cross-segment pair silently skips.

The cross_seg_ack gate is the good idea here

The pack has two bridging routes, and prev_tail means different things in each. Under the copy bridge, the pinned frames are a bit-exact copy of the previous segment's tail - so prev_tail is the real reference. Under the conditioning-only route, those frames are this segment's own redraw, so prev_tail is only an approximation.

The author measured the proxy error at 0.006, against a seam step of 0.0007 that he was trying to fix. Aligning to a worse reference than the error you're correcting doesn't help - it made the first frame worse. So cross_seg_ack defaults to false, and match_prev/lowfreq_pull abstain rather than act on a bad reference. Tick it when your guide is genuinely the previous segment's last frame. The baseline field is the same instinct applied to the within-segment work: robust (default) computes dispersion and lets groups abstain if the reference isn't trustworthy, rather than moving pixels on a bad baseline.

Install

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

ComfyUI Manager → search h3-latent-relay, restart the backend. No models, no extra packages - this node is pure image maths on your existing install, which is part of why it's the cheapest of the optional four.

Traps

Keep every dial at 0 until you've looked at the seam and can name what's wrong with it. Enabling match_prev and lowfreq_pull together is discouraged - their scopes overlap, pick one. And if hist_match feels slow, it's the high-resolution subsampling path kicking in above roughly 0.8 MP; it's a fixed couple of seconds per segment, not proportional to resolution. Paying for a fix you can't see is worse than the seam.