Atlas Fill Occluded 🕳️‍🩹

The moment your camera moves off the recovered viewpoint, it reveals parts of the scene the photograph never saw - the far side of a column, the ground behind a foreground figure. Those are disocclusion holes, and they're the thing that makes a 2.5D projection read as a cheap card trick instead of a move. Most fill approaches answer the wrong question: they fill a hole in a render, which is only valid for the exact camera path it was generated against. Change the angle by a degree and the fill is worthless, and nothing ever comes back into the scene.

Atlas Fill Occluded 🕳️🩹 fills a hole in the scene. Each fill is registered back as a ProjectionSource - real geometry the solve owns - so it exports to your DCC and every later camera move reuses it instead of re-inventing it. That's the whole architectural difference, and it's why this node sits in Atlas/advanced and why the fix it does actually sticks.

How it works

For each move-revealed cluster of holes it runs a per-ROI loop: survey the hole regions → crop the ROI at its own native raster → fill it with an edit model at that size → composite the fill back into the plate → register the fill plus its crop camera back as geometry via AtlasAddPatchView. It fills at most max_rois clusters (default 8), largest first, with min_area_px (default 1024) as the floor. That budget is on purpose: leftover speckle at the rim of a tear is cheaper handled by a band layer than by a diffusion pass each. Residual hole after a defaults run is expected - it isn't evidence the node failed.

The fill model defaults to flux_fill, wired through InpaintModelConditioning so the noise mask confines the sampler to the hole. qwen_edit is the alternative (Qwen Image Edit with an inpainting LoRA is the author's measured pairing for registering fills). If you pick flux_fill, remember its guidance wants to be high - the default flux_guidance of 30 is there because Fill models are not ordinary FLUX. The node takes model, clip, vae sockets like any loader-based graph, and fill_model is not validated against what you wire in - matching them is on you.

What matters most

• camera_path - the move whose disocclusion you're filling. Holes are surveyed at its final frame, whose holes are a superset of every earlier frame's. No path = the solved pose, which by definition has almost nothing to fill.
• primary_depth - metric depth of the primary scene. The register_to_primary camera mode requires it; without it every ROI falls to the failure policy.
• prompt - a sensible default ("continue the surrounding surfaces, do not invent new objects") that reads like it was written by someone who's seen a fill model wallpaper a hole with a beach. min_gen_long_edge exists for the same reason: below ~1024, small clusters reach FLUX at native size and answer with nonsense - the author measured two 256×256 ROIs both answered with sand.

Outputs are the repaired image, the updated solve, and a report. Read the report's fillable hole numbers, not the raw peak hole figure - the raw number counts sky and off-plate pixels and will make a good run look like a failure.

Installing

Part of atlas-camera:

cd <ComfyUI>/custom_nodes
git clone https://github.com/mikejamesvfx/atlas-camera.git

Restart. The heavy lifting is model loading, which you already know how to do - you need FLUX Fill (or Qwen Image Edit) plus CLIP and VAE on the inputs. No non-commercial model is required by the pack itself.

Honest expectations

This is the most demanding node of the batch, and the author's own notes are refreshingly blunt about what "worked" means: at defaults on their test shot, fillable hole went from 4.56% to 1.28% of frame - 72% of the real disocclusion closed, zero pixels opened. That's the shape of success: not zero holes, but most of the real ones gone with none invented. And the default camera_source (declared_orbit) constructs the end pose exactly rather than measuring it, because a measured pose of a generated view was demonstrably worse than the constructed one. If you switch to register_to_primary, on_registration_failure decides what happens to ROIs the registration won't stand behind - skip for unattended loops, declared_orbit for artist-mode fallback. For a first run, leave the defaults alone and read the report.