Masked Editing

Masked editing

Choose the operation

Goal	Start from
Inpaint, remove, replace, or outpaint a region	Masked Video Inpaint
Continue a clip with protected picture and sound	Masked AV Extension — Single Clip
Continue several reviewed scenes	Masked AV Extension — Chain + Reference Image
Preserve both endpoints and generate the gap	Two-Clip Masked AV Bridge

Spatial editing path

1. Loop Source AV Target selects and encodes the current source interval.
2. Loop Mask Slice broadcasts one mask or selects the matching tracked-mask frames, including the repeated continuation overlap.
3. Grid Preview shows the effective H3 generation region.
4. Apply Target Mask attaches the spatial and audio denoise masks to the real joint H3 target.
5. The normal loop handles sampling, checkpoints, review, resume, and assembly.

What the mask means

• white = generate is the conventional inpainting view.
• white = preserve is useful when the supplied artwork describes protected content.

Use the same convention on Grid Preview and Apply Target Mask.

H3 reads effective 32×32 source-pixel cells. Exact conversion follows the video VAE's causal 1,4,4,4,4 source-frame groups and its 2×2 latent-token cells. A tracked mask must contain the exact raw scene span; a one-frame mask is broadcast. Legacy trilinear conversion remains available only for comparing older renders.

Ordinary AV extension creates its own temporal prefix mask and does not need a user-supplied spatial mask.

Audio mask choices

• Preserve source audio: the complete target audio stream is protected.
• Generate audio: the target audio stream remains denoisable.
• Custom audio mask: supply a time mask when only selected sound regions should regenerate.

Video and audio masks use 0 = preserve, 1 = generate after the user's mask meaning has been resolved.

Composition with chain continuation

Apply Target Mask intersects an existing nested AV mask rather than replacing it. A protected incoming prefix therefore stays protected while a spatial mask selects additional generation cells in the remainder of the scene.

Outpainting

Prepare source frames at the final H3 canvas size before encoding. Place the source image in the region to preserve, fill the expanded canvas as intended, and provide a mask whose generation region covers the new area. Width and height must remain divisible by 32.

Two-ended bridge

Two-Clip AV Bridge places a source tail and destination head into one empty joint target. Both endpoint windows are protected; only the middle is denoised. A 39-frame endpoint maps exactly to 12 video and 65 audio latent steps. The bridge is a single target operation rather than a recursive loop.

Changing mask-conversion mode changes generation. Update the Plan generation fingerprint and regenerate affected accepted scenes before resuming.

Next: Node reference, Continuity and audio, or Troubleshooting.