ComfyUI Node
Atlas Clean Plate Layer 🖼
A ComfyUI node in Atlas/07 · Clean Plate & Inpaint with 40 inputs and 4 outputs.
Atlas Clean Plate Layer 🖼
- solve
- depth
- plate_image
- plate_ref
- exclude_mask
- layer_matte
- band_split
- band_ref_mask
- plate_depth
- object_mask
- solve
- hole_mask
- extend_mask
- report
◄near_m0.0►
◄far_m0.0►
◄near_pct0.00►
◄far_pct0.50►
◄namelayer►
◄priority0►
◄relief_grid384►
◄depth_edge_rel1.50►
◄fill_occludedfalse►
◄embed_mattefalse►
◄edge_extend_px0►
◄skirt_bevel0.00►
◄frame_outpaint_px0►
◄exclude_choke_cells2►
◄band_sidemanual►
◄band_geometryrelief►
◄geometry_override►
◄band_override►
◄max_edge_factor12►
◄normal_edge_deg0►
◄quad_coherencetrue►
◄sub_quad_boundaryfalse►
◄silhouette_mattefalse►
◄soft_visibilityfalse►
◄transition_ribbonfalse►
◄ribbon_px64►
◄ribbon_bend-0.30►
◄ribbon_adaptivefalse►
◄ribbon_depth_slope2.00►
◄ribbon_smudge_px12►
CategoryAtlas/07 · Clean Plate & Inpaint
Inputs (40)
| Name | Type | Default | Description |
|---|---|---|---|
| solve | ATLAS_SOLVE | — | |
| depth | ATLAS_DEPTH_MAP | — | |
| plate_image | IMAGE | — | |
| near_mopt | FLOAT | 0.00–10000 | MUST match the AtlasDepthLayerMask band that produced plate_image. |
| far_mopt | FLOAT | 0.00–10000 | MUST match the AtlasDepthLayerMask band that produced plate_image. |
| near_pctopt | FLOAT | 0.000–1 | Must resolve to the same band as the AtlasDepthLayerMask that produced plate_image (both call the shared _resolve_depth_band helper, so identical near_m/far_m/near_pct/far_pct here and there always agree). LOWER near_pct = tighter occlusion, not looser — see AtlasDepthLayerMask's near_pct tooltip for the worked example. |
| far_pctopt | FLOAT | 0.500–1 | Must resolve to the same band as the AtlasDepthLayerMask that produced plate_image. 0 means no upper bound (+inf). |
| nameopt | STRING | layer | — |
| priorityopt | FLOAT | 00–100 | Blend priority among layers (higher wins). FARTHEST bands take the HIGHER priority (far 15 / bg 10 / mid 5 / fg 0): at a watertight seam the surfaces are depth-adjacent and this near-tie bias picks the winner, so nearest-highest renders a band's edge smear IN FRONT of the layer behind it (striped seams). Min is 0; a sky dome goes negative via AtlasSkyDomeLayer. Ordering is by depth + priority, never facing angle (clean-plate sources paint head-on AND grazing, unlike multi-angle patches). |
| plate_refopt | ATLAS_PLATE_REF | Optional registered final clean-plate reference. Browser still uses image_b64 preview; exporters use this for EXR/float-safe handoff. | |
| relief_gridopt | INT | 38416–4096 | Band-clipped meshes tear at band boundaries ON TOP OF normal silhouette tearing, so per-layer meshes want more density than the generic 128 default - 384/1.5 is the empirically-calibrated band-layer setting (hangar + monument valley). |
| depth_edge_relopt | FLOAT | 1.500.05–5 | Looser than the generic 0.5: safe WITHIN a band because the band clip already bounds the mesh's depth range. |
| exclude_maskopt | MASK | Optional external exclusion (e.g. a real sky segmentation from SAM/RMBG). When connected it REPLACES the internal sky heuristic (which otherwise eats tall geometry above the horizon) - should match whatever was passed to the AtlasDepthLayerMask call that produced plate_image, for hole_mask/band resolution to stay in lockstep. | |
| fill_occludedopt | BOOLEAN | false | Diffusion-fill this band's mesh across the foreground occluder's footprint (the band clip otherwise leaves a hole exactly there) so the INPAINTED plate content lands on real geometry instead of a hole - the disocclusion 'shadow ray' mesh. Synthesized depth is a smooth interpolation of the surrounding background, reported in metadata as n_filled_cells. Excluded (sky) regions are never filled. |
| embed_matteopt | BOOLEAN | false | Embed a full-resolution per-pixel edge matte on this layer (ProjectionSource.mask_b64) - the projection shader then cuts the TRUE band silhouette per-pixel instead of the mesh's blocky grid-resolution tear edge, and the Nuke layers export writes it into the plate's alpha. Auto-computed from this band (in-band pixels, plus the filled occluder footprint when fill_occluded is on, minus exclude_mask); wire layer_matte to override with a hand/SAM matte. |
| layer_matteopt | MASK | Optional explicit edge matte (overrides the auto-computed band matte when embed_matte is on) - e.g. a SAM segmentation of this layer's subject for a crisper edge than depth banding gives. | |
| edge_extend_pxopt | INT | 00–512 | Deterministic edge-extend for THIS layer, same trick as the sky dome's: smear the plate's colors past the matte edge by this many pixels (quarter-res neighbor propagation — NOT an inpaint), dilate the embedded matte to expose the extension on disocclusion, and grow the mesh's boundary skirt to receive it. The invented region is reported on the extend_mask output AND exported to Nuke/Maya as {layer}_extend_matte.png so it can be processed downstream (regrain, blur, replace). Smeared pixels are plausible only for narrow slivers — large reveals still want a real inpainted plate. Turns on embed_matte implicitly (an extension needs a matte edge). |
| skirt_bevelopt | FLOAT | 0.000–4 | Bevel the mesh's boundary skirt AWAY from the camera, as a slope in local cell units: 1.0 recedes one cell per extension ring (a 45° skirt), 0 = today's flat skirt. Physically motivated — an occluded surface continues away from the camera behind its silhouette, so a receding bevel is the least-wrong geometry at a tear edge. Try 1.0–2.0 with edge_extend_px: the smeared colors land on the receding skirt and the extend matte marks them for regrain. |
| frame_outpaint_pxopt | INT | 00–1024 | Outpaint THIS layer past the FRAME edges by this many pixels — the same per-source widened-camera trick the sky dome uses: the plate canvas is padded edge-replicated, this source gets its OWN camera with shifted cx/cy and grown W/H (the primary solve and every other layer are untouched), and the band mesh extends past the original frustum to carry it. Closes the frame-edge reveal that 🧭 Safe Zone measurements show is the binding constraint on wide scenes (ground layers used to end exactly at the photo boundary). The ring is INVENTED pixels: it lands in extend_mask / {layer}_extend_matte.png for downstream regrain, and turns on embed_matte implicitly. 0 = off. |
| exclude_choke_cellsopt | INT | 20–16 | Choke-and-reskirt against the exclude_mask edge: segmentation and depth edges never align exactly, leaving a ribbon of cells the mask calls rock but whose depth IS sky — they back-project high above the real silhouette as a jagged floating band. This erodes the layer N grid cells away from the exclusion, then the boundary skirt regrows the ring with clean neighbor depth: same coverage, geometry hugging the true surface. Raise for sloppier segmentation masks; 0 disables. |
| band_sideopt | COMBO | manual | With band_split connected: foreground = [0, split), background = [split, +inf) — the node's own near/far widgets are ignored. manual = use this node's own near/far settings. |
| band_splitopt | ATLAS_BAND_SPLIT | Wire ONE AtlasDepthBandSplit into every band node (with band_side set) so the fg/bg boundary lives in exactly one widget and the layers can never drift apart. | |
| band_geometryopt | COMBO | relief | How this band's projection surface is built. relief (default) = the depth-following mesh, for anything with real 3D shape inside the band. card = ONE flat fronto-parallel plane at the band's median depth (classic DMP card) — for distant/flat-facing layers with negligible internal parallax (far mountains at the horizon, a hangar's back wall, a skyline backdrop); never tears, zero depth noise. ground = the exact analytic Y=0 ground plane — for flat horizontal surfaces the camera stands over (desert floor, water, road); zero depth-noise bumps. Both flat modes keep band membership from the REAL depth (which pixels belong) and only flatten WHERE they sit; matte/edge-extend/outpaint all still apply. |
| geometry_overrideopt | STRING | Optional geometry-type override STRING — wins over band_geometry when non-empty ('relief'/'card'/'ground'). Exists because ComfyUI rejects STRING→combo links: wire AtlasAssessImage's geom_far/bg/mid/fg output here so the VLM's per-layer geometry recommendation flows in (same pattern as patch_view_override). Unknown values error loudly. | |
| band_ref_maskopt | MASK | Exclusion used ONLY for resolving near/far percentages to metres. When exclude_mask carries per-layer scoping (🎯 scope rows), each layer's depth population differs and the shared band edges DRIFT apart (metric gaps between adjacent bands — debug-report finding). Wire the plain SKY mask here on every band node so all layers resolve identical edges. Unwired = legacy behavior (band edges from exclude_mask's population). | |
| band_overrideopt | STRING | Optional band override STRING ('near_pct=<f> far_pct=<f>') — wins over this node's near/far widgets when non-empty. Wire AtlasAssessImage's band_far/bg/mid/fg output here so the VLM's subject-aware band boundaries flow in (jointly derived, so adjacent bands always share edges exactly). MUST be the same string the paired AtlasDepthLayerMask received. Loses to a connected band_split. Errors loudly on garbage. | |
| max_edge_factoropt | FLOAT | 122–200 | World-space edge tear threshold (SEPARATE from depth_edge_rel). Dominant tear cause on deep / narrow-FOV / interior bands: raise to 40-80 to stop comb-tearing continuous grazing surfaces. >80 rubber-sheets real silhouettes. |
| normal_edge_degopt | FLOAT | 00–180 | 0 = off. Tears where surface NORMALS bend past this angle — real creases / occlusion silhouettes — while leaving smoothly-receding walls intact. Pair with a higher max_edge_factor: raise mef to kill spurious combs, then ~40-70 here to keep genuine edges torn. |
| quad_coherenceopt | BOOLEAN | true | Reject both triangles when either half of a grid quad fails; avoids surviving diagonal UV wedges at band boundaries. |
| sub_quad_boundaryopt | BOOLEAN | false | Cut a torn cell AT the depth cliff instead of deleting the whole cell. A band silhouette can otherwise only turn in whole-cell steps AND loses a cell of real surface on both sides of every cliff - measured 5.67px mean boundary error at grid 128 on a 1024px plate (step 8px), WORSE than the 4px quantization bound. Reads the cliff from the full-resolution depth and rebuilds both sides up to it, never joining them: 5.67 -> 1.43px, camera coverage 97.8 -> 100%, ~5% more vertices. Same thresholds, same cells torn. |
| silhouette_matteopt | BOOLEAN | false | Cut this band's exclusion silhouette per-pixel rather than at grid resolution: grows a boundary skirt and ships a full-resolution matte the projection shader trims it back with (strip uncovered 0.083 -> 0.000, zero pixels spilled past the boundary, edge within 0.25px of truth vs 4px lattice). Skirt and matte are ONE switch - an unmatted skirt shows replicated pixels. Handles the sky/exclusion edge; depth cliffs need sub_quad_boundary. |
| soft_visibilityopt | BOOLEAN | false | SOFT LAYERING (SLIDE, ICCV 2021) - the viewport answer to sawtooth silhouettes. Instead of TEARING at a depth cliff and leaving a hole to feather, keep the surface continuous (it rubber-bands across the cliff) and fade those fragments with a per-pixel visibility A = exp(-beta*|grad disparity|^2) computed at PLATE resolution. No tear means no boundary to quantize, so there is no staircase at ANY grid - and the lattice stays intact, so planar hole patch and the CUDA repair keep working (unlike sub_quad_boundary, which this supersedes and disables). Measured on a diagonal cliff: fade to 0.039 over a 4px feather while smooth receding ground stays at 0.987. Needs something BEHIND to reveal (BG clean-plate / sky dome / clean_plate input) or the fade shows the backdrop. Export still tears - a DCC has no shader to fade with. |
| transition_ribbonopt | BOOLEAN | false | KEEP this band's tear and hang a bounded skirt off it, instead of soft_visibility's untorn rubber band whose fins run as far as the depth jump dictates. Separate topology grown from the open rim, stepping outward in IMAGE space so its apparent width is a fixed pixel count at any depth (a view-ray extrusion cannot do this - every point on a ray projects to the same pixel). UVs freeze at the silhouette texel, so it is an edge-extend CLAMP, which is what the seam doctrine asks of a layer that has bands behind it. Shares no vertex with the band surface. Mutually exclusive with soft_visibility. |
| ribbon_pxopt | FLOAT | 640–400 | Transition-ribbon width in SCREEN pixels of the plate, measured in the recovered camera; achieved width is reported back as stats.transition_ribbon.measured_px_p50/p95. Distinct from edge_extend_px, which smears TEXTURE outward on this band; this adds GEOMETRY carrying the frozen rim texel. |
| ribbon_bendopt | FLOAT | -0.30-0.5–0.5 | Shape of the depth falloff toward the band behind. NEGATIVE curls away fast then levels off (the tight lip); 0 is linear; POSITIVE dwells at this band's depth then dives (a flange). Limited to +/-0.5, the range in which the ramp stays monotonic. |
| ribbon_adaptiveopt | BOOLEAN | false | Scale ribbon_px by the size of the discontinuity relative to depth_edge_rel, clamped 0.5x-2x. DEFAULTS OFF: measured on a real exterior it was the only setting that broke width consistency (p95 85px for a requested 64) and clamped the most folds. |
| ribbon_depth_slopeopt | FLOAT | 2.000.25–8 | How far the skirt may recede toward the band behind, as a multiple of its own world width. Governs what an ORBIT sees; ribbon_px alone bounds only the recovered camera. Low (0.5-1) hugs this band's silhouette, high (4-8) reaches further back at the cost of long slabs off-axis. |
| ribbon_smudge_pxopt | FLOAT | 120–200 | Blur the skirt along this band's silhouette, in plate texels, ramped in from the rim. At 0 each column is one frozen texel and the skirt reads as flat radial streaks. Applies in the viewport and is baked into the exported GLB as vertex colour. |
| plate_depthopt | ATLAS_DEPTH_MAP | AtlasDepthMap run on plate_image (MoGe, same solve wired). When connected, THIS layer's mesh is built from the clean plate's depth instead of `depth`: the terrain behind the foreground object comes from the plate that actually sees it. Its scale is registered to `depth` by median ratio over the mutually visible pixels (outside object_mask); the report carries scale + rel_mad. Same camera, so no second solve and nothing to line up. | |
| object_maskopt | MASK | The foreground object (white) in the SOURCE frame — excluded from the scale registration (its depth differs between the plates by definition). Recommended when the object is large. |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| solve | ATLAS_SOLVE | — |
| hole_mask | MASK | — |
| extend_mask | MASK | — |
| report | STRING | — |