Chaining H3 clips without the seams: latent slicing and audio that actually continues

What it does

It chains MiniMax H3 generations into one continuous piece: clip B gets the tail of clip A pinned to its head, so motion, speed, direction, and the soundtrack carry on across the join instead of restarting. The trick is that the pinned frames are sliced directly out of the previous clip’s latent — the same numbers the model produced, no decode/resize/re-encode round trip. Audio is treated as a first-class citizen: the pinned window ends at the join and reaches backwards into what already played, so the model continues your track rather than improvising a soundalike.

The interesting bit

The audio half is the genuinely novel part. ComfyUI’s stock Add Guide anchors audio forward from a frame, which produces a new take that sounds similar; this pack’s backwards-reaching window is what makes a beat stay on beat. Also notable: the nodes verify their assumptions against ComfyUI’s live layout code at first use and refuse to run if an update moved something — a loud failure instead of a silently bad render.

Key highlights

• Latent-to-latent chaining: no pixel round trip, so no color drift or softening down long chains
• Audio continuation with a real seam check (the Seam Probe measures whether a join is a true continuation)
• Save/Load Latent pair carries clips across runs; files also travel through ComfyUI’s HTTP API for remote tooling
• A Chain node handles the approve/re-roll loop, including auto-advancing clip indices
• Plays nicely with Add Guide, reference mode, and last-frame targets — with explicit warnings when anchors conflict

Caveats

• Requires ComfyUI 0.34.0+ (arbitrary keyframe anchors); older versions need the 0.3.1 release
• Resolution must match between clips — latents can’t be resized, and the node refuses rather than silently degrading
• Some treble loss at every audio join; longer audio windows get progressively muffled and quieter
• The Chain node only sees Load/Save if all three sit in the same canvas group — an easy thing to miss

Verdict

Anyone building long-form H3 video with sound — music especially — should try this; it solves two problems nothing else in the ecosystem addresses. If you only need to anchor a still or short clip, stock Add Guide already does that and you don’t need the pack.