Continuity and Audio

Continuity and audio

Two policy nodes answer separate questions:

1. Transition Policy: what enters this scene from its predecessor?
2. Audio Policy: what guides generation, continues between scenes, and becomes the final soundtrack?

Choose a transition

Transition	Carries into the scene	Recommended use
Cut	Nothing	Start an independent scene.
Guide	22 clean RGB/VAE frames	Default motion handoff. Start here.
Tone Carry Guide	Guide plus a detected tone correction	Experimental tone-continuity comparison.
Latent Guide	22 sampled video-latent guide frames	Opt-in path that avoids an RGB/VAE round trip.
Detail Guide	Disposable tapered chroma-noise Guide	Experimental identity/detail comparison.
Hard AV	Exact protected 39-frame picture prefix and optional sound	Strict AV-prefix continuity.
Soft AV	Exact picture plus an eight-tick generated-audio release	Recommended AV-mask handoff.
Detail AV	Disposable tapered-noise Hard AV picture prefix	Experimental 39-frame latent-noise comparison.

The choice describes the boundary entering a scene. It does not redefine the completed predecessor.

AV-mask transitions require video encode, head anchor, and an exact shared video/audio boundary. Start with 39 context frames. Expert shared-clock values are 39, 90, 141, 192, or 243 frames; Detail AV v2 specifically requires 39.

Choose audio behavior

Decision	Choices	Meaning
Final audio	Generated / Source / None	What Assemble places in the MP4.
Source reference	On / Off	Whether the current source window guides H3.
Generated continuity	On / Off	Whether the preceding sampled audio latent enters the next scene.

Common combinations:

Goal	Final	Source reference	Generated continuity
Let H3 create a continuous soundtrack	Generated	Off	On
Preserve a song, narration, or dialogue track exactly	Source	On	Off
Preserve the source track but also carry generated sound internally	Source	On	On
Produce silent final video	None	As needed	As needed

Generated scene audio is still saved as WAV sidecars when connected to Segment

• Checkpoint, even if Source becomes the final soundtrack.

Trimming and final timing

Loop Trim removes the repeated picture prefix. With match_tail=true, it also time-conforms the small H3 audio-grid mismatch to the exact delivered-picture duration. Connect its AUDIO output directly; masked-AV overlap audio travels privately inside that output to Segment + Checkpoint.

images_with_overlap is a separate visual stream for assembly blending. It does not alter the normal hard-trimmed images or audio.

Generated audio sidecars

When decoded audio reaches Segment + Checkpoint, each accepted scene saves an uncompressed WAV under:

ComfyUI/output/h3_chains/<run_name>/generated_audio/

Completed assembly also writes a combined generated WAV even when Source is the selected final soundtrack. This keeps H3 ambience, effects, or performance available for post-production.

Experimental source-reference alignment

For a 362-frame source window, stock H3 can produce one extra audio-reference latent step after padding. Current Shot's align_audio_reference option trims only the Ref2VA slice to 15.070 seconds. It does not change the full source track used by assembly. Leave it off for stock frame-exact comparisons.

Measure an audio join

Place Seam Probe between the current untrimmed audio decode and Loop Trim. Connect the predecessor AV latent, H3 audio VAE, and the same trim_frames. It reports timing offset, correlation, broadband level step, and low-frequency ambience-floor change. Strongly periodic music can alias by one full cycle.

Next: Source Timeline and tagged audio, Plan reference, or Troubleshooting.