Atlas Crop ROI ✂️

When a camera move opens a hole in your projection - a disocclusion gap where the source photo never looked - the fix is to generate the missing pixels with a video or image model. But you don't want to generate the whole frame; you want to aim the generator at exactly the hole. Atlas Crop ROI ✂️ is the node that turns one artist-drawn fill region into a generation-ready crop plus the camera that region was rendered with.

The mechanism

The key insight is that a camera-matrix crop is just a shifted principal point with a smaller raster - no separate render, no perspective warping, no misalignment. The ROI renders 1:1 with the plate through the same rasterizer the solve uses, so what comes out is geometrically exact. That's why the paste-back later (via AtlasCompositeCrop) is a pure rect blit instead of an alignment exercise.

The node handles the variable-ROI problem with an artist budget: three slots (roi_slot 1–3), three static branches, and an unused slot degrades to a no-op end to end - its guide passes through, the composite returns its input. That "no-op when empty" behavior is what lets you build one static graph that works whether you drew one, two, or three fills.

What comes out

The outputs are generation-ready on purpose:

• guide - the crop's image, aspect-preserving and snapped to a /16 grid, no longer than max_gen_long_edge (1280 default). A raster your generator will actually accept without resizing surprises.
• mask - the hole mask for that crop, at the same raster.
• gen_width / gen_height - the raster as plain INTs, made to wire into WanVaceToVideo's width/height widgets (which would otherwise hard-code one raster for every crop).
• crop - the ATLAS_CROP handle carrying the native rect for the exact paste-back.

Artist or automatic?

roi_source defaults to artist: the viewport's drawn Fill ROIs win. But auto_largest is worth knowing - it surveys the end frame's disocclusion holes and takes the roi_slot-th LARGEST cluster, which is exactly the holes the viewport wand can't fix. It uses the same clustering recipe as the pack's CLI, so the two can't drift. min_area_px filters out clusters small enough for the wand or a band layer to handle, and exclude_mask subtracts regions already carried by something other than geometry (a sky dome, a matte) so the generator is never aimed at them. Auto mode also always drops holes whose ray at infinity lands on a hole in the plate (sky - nothing was ever occluding it) or outside the plate frame entirely (that's outpainting, not filling).

The camera_path input matters for the move case: give it the move and the crop renders its FINAL frame - whose holes are a superset of every earlier frame's - so one fill covers the whole path. No path means the solved pose.

Other knobs: pad_frac (context around the rect, 0.1 default), snap (grid snap, 64), hole_dilate_px (4). Install is the pack's clone-and-go; no heavy extras for the crop math itself. It's filed under Atlas/advanced because it's part of the repair loop, but it's honestly the friendliest way to aim a generator at a hole.