Skip to content

av-latent-extend.md

Latest commit

 

History

History
54 lines (44 loc) · 2.61 KB

av-latent-extend.md

File metadata and controls

54 lines (44 loc) · 2.61 KB
Raw

😺NKD AV Latent Extend

Continues a MiniMax H3 clip into a new one, sound included. The tail of the segment you already sampled is planted at the head of a fresh empty latent and masked as "keep", so the sampler doesn't start a new shot: it carries on the motion and the sound that were already there. Chain it once per stage to build a long take out of short ones.

flowchart LR
    PREV(["KSampler<br/>(previous stage)"]):::external -- previous --> EXT
    NEW(["MiniMax H3<br/>Reference To Video"]):::external -- "new (empty)" --> EXT
    EXT["**NKD AV Latent Extend**"]:::nkd -- latent --> KS(["KSampler"]):::external
    EXT -- overlap_frames --> IBE(["Image Batch Extend<br/>With Overlap"]):::external
    EXT -- trim_seconds --> TAD(["Trim Audio Duration"]):::external

    classDef nkd fill:#3b3b6b,stroke:#8ab4ff,stroke-width:2px,color:#fff
    classDef external fill:#2d2d2d,stroke:#888,color:#eee
Loading
  • previous is the previous stage's KSampler output, before decoding. It needs the latent, not the pixels.
  • new (empty) is the empty AV latent of the next segment, straight from MiniMax H3 Reference To Video. Its length is what gets rendered, and the overlap eats its first chunks.
  • Overlap Chunks (default 2) is how much of the previous clip the model sees to continue from. One chunk is 5 frames and each extra one adds 17, so 2 chunks is 22 frames, about 0.9 s. More overlap buys smoother continuity and costs new footage per stage.

Three outputs, and the last two exist because the overlap has to come back off after decoding:

  • latent goes to the KSampler.
  • overlap_frames feeds Image Batch Extend With Overlap (side=source, mode=cut) to drop the regenerated head from the picture.
  • trim_seconds feeds Trim Audio Duration's start_index. It is the same overlap measured on the audio latent grid (1/40 s). Cutting the sound at the video frame time instead drifts a few milliseconds every stage, and those add up over a long chain.

The seam is hard on purpose. Nothing is blended or faded: the planted tail is pinned at "keep" and the model does the mixing itself over the following frames, which is what makes the continuation move like the clip it came from rather than dissolve into it. The overlap is regenerated context, so it is thrown away in pixel space afterwards, and only what comes after it survives.

It refuses rather than guesses: a previous and a new latent of different sizes, a previous clip shorter than the overlap you asked for, or a new latent no longer than the overlap all stop with a message saying which.

← All 😺NKD Basic Tools nodes