[Legacy] H3 Continuum Assemble + Seam

If you've generated a chained video longer than the model's native limit, you know exactly where the magic dies: the boundary between chunks. One frame goes bright, there's a micro-flash, or the audio clicks. H3 Continuum Assemble + Seam is the current V3.3 assembler from the H3 Continuum pack, and it exists to make those boundaries invisible instead of making you pray.

It's the second half of the standard Continuum workflow. The sampler gives you raw video_latents and audio_latents per chunk; you decode those with normal ComfyUI Core VAE Decode nodes; then everything lands here. The node trims the decoded overlap (each chunk carries continuation context, so its first frames duplicate the end of the previous chunk), re-aligns audio on cumulative frame boundaries, and applies two guarded correction passes.

The inputs that matter

• images - the list of decoded chunk images from Core VAE Decode. Full raw chunks get decoded before any trimming; trimming the latents pre-decode would break temporal VAE conditioning, so don't do that yourself.
• audio - the decoded chunk audio.
• assembly_plan - straight from H3 Continuum Sampler V3.3's assembly_plan output.
• exact_total_duration (true) - trims or minimally pads the final output to exactly chunks × seconds × 24fps.
• audio_seam (Auto) - corrects decoded audio boundaries only. It never touches video frames.
• video_seam (Auto) - the headline feature. Auto applies the validated transient and micro-flash correction with no frame deletion; Auto 2 additionally enables guarded exposure-ramp smoothing (marked experimental); Analyze Only reports the boundary classification without changing anything; Off skips it.

The "guarded" in the description is the whole design philosophy: video correction is only adopted when the boundary and its five-frame transition clearly improve - the pack would rather leave a small seam than smear a whole scene with a naive blend. The author's validation notes call Auto 2 the fallback when Auto doesn't tame a visible exposure ramp at a boundary.

Outputs are images, audio (both ready for a Create Video / save node), and a report STRING you can drop into Core Preview as Text to read exactly what each boundary was classified as and what was applied.

Install and troubleshoot

cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git

or install "H3 Continuum" via ComfyUI Manager, then restart. No extra pip packages - just a working MiniMax H3 ComfyUI environment and ComfyUI v0.32.0+ for the current Core VAE fixes.

When seams still show: run once with video_seam = Analyze Only and read the report to see what the classifier found at each boundary - that tells you whether it's a transient/flash (Auto handles it) or an exposure ramp (try Auto 2). If the whole boundary looks like a hard cut rather than a flash, the real fix is usually upstream: more continuity frames on the sampler or a better Timeline prompt that keeps action/camera continuous across the chunk break. And remember the pack's own caveat - none of this guarantees frame-perfect continuity; it just removes the artifact the model itself didn't cause.