ComfyUI Node
Atlas Input 🎬
A ComfyUI node in Atlas/01 · Input & Camera with 29 inputs and 5 outputs.
Atlas Input 🎬
- image
- raw_meta
- solve
- image
- depth
- sky_mask
- report
◄layers0►
◄meshrelief►
◄mesh_resolution512►
◄use_vlmfalse►
◄vlm_providerlmstudio►
◄vlm_model►
◄skyfalse►
◄sky_promptsky►
◄scope_prompts►
◄inpaintfalse►
◄upscale_model►
◄edge_extend_px24►
◄max_edge_factor12►
◄sky_heuristictrue►
◄normal_edge_deg0►
◄depth_modeldepth-anything/Depth-Anything-V2-Metric-Outdoor-Large-hf►
◄vlm_scopetrue►
◄sky_inpaint_modelama►
◄sky_lama_grow_px32►
◄sky_sdxl_checkpointSDXL/sd_xl_base_1.0.safetensors►
◄sky_sdxl_positiveclear seamless sky, high detail, no buildings, no trees, no roofs►
◄sky_sdxl_negativebuilding, tree, roof, person, vehicle, text, watermark, blurry►
◄sky_sdxl_seed0►
◄retopo_methodoff►
◄retopo_target_vertex_count2000►
◄boundary_smooth_iterations0►
◄sub_quad_boundaryfalse►
CategoryAtlas/01 · Input & Camera
Inputs (29)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | — | |
| layersopt | INT | 00–4 | 0 = one full-range mesh (instant relief). 2/3/4 = depth-band clean-plate layers on the proven splits (2→0.55; 3→0.2/0.65; 4→0.3/0.6/0.8), watertight by construction. 1 = one full-range clean-plate layer (useful with mesh=card/ground). |
| meshopt | COMBO | relief | layers=0: relief = depth-following mesh; card/ground = ONE flat plane (band-median card / analytic ground). layers>0: the DEFAULT band geometry — the VLM's per-band call wins when use_vlm is on. |
| mesh_resolutionopt | INT | 512128–2048 | Relief grid (long-edge cells). Internal tear threshold pairs automatically: 0.5 for the single full-range mesh, 1.5 for band-clipped layers (the calibrated pairings). |
| use_vlmopt | BOOLEAN | false | Run the 🧭 VLM assessment first (advisory — never blocks; VRAM offloaded after) and wire its SAM prompts, per-band geometry, and band boundaries into the inner nodes. With layers>0 this forces the 4-band plan (the VLM's plan has 5 fixed slots). |
| vlm_provideropt | COMBO | lmstudio | 4 options: ollama, lmstudio, llamacpp, openai |
| vlm_modelopt | STRING | Blank = the provider's default model. | |
| skyopt | BOOLEAN | false | SAM-segment the sky onto its own flat card, and feed the mask into every mesh's exclude_mask + band_ref_mask. Uses native SAM3 (transformers>=5.5.4, [sam3] extra) or falls back to AtlasSemanticMask — skipped + noted if neither is available. |
| sky_promptopt | STRING | sky | Manual sky segmentation prompt; the VLM's wins when use_vlm. |
| scope_promptsopt | STRING | Manual per-band SAM scoping, ONE PROMPT PER LINE far→near (line 1 = farthest band). Blank line = that band stays band-only. Self-disarming: a no-match segment falls back to band-only automatically. The VLM's prompts win when use_vlm. Uses native SAM3 ([sam3] extra) or AtlasSemanticMask. | |
| inpaintopt | BOOLEAN | false | Build each occluded band's clean plate: occlusion mask → expand → ✂crop → LaMa → ✂stitch (the 256²-bottleneck fix). Needs comfyui-inpaint-nodes + big-lama.pt — skipped + noted if absent. Off = bands project the original photo (honest holes on reveal). |
| upscale_modelopt | STRING | Optional upscale model FILENAME (models/upscale_models) fed to the inner LaMa nodes — e.g. 4xRealWebPhoto_v4_dat2.safetensors. Measured 6.5× fill detail vs legacy. Blank = off. | |
| edge_extend_pxopt | INT | 240–256 | Behind-band edge-extend (layers>0): how far plate colours smear PAST each silhouette to hide grid-step tears — the frontmost band always keeps a clean 0 cut. Was baked at 64 (tuned for smooth ridgelines); lower it for high-frequency content like foliage, which 64 shreds into halos. 0 = clean cut on every band. |
| max_edge_factoropt | FLOAT | 122–200 | World-space edge tear threshold (layers=0 relief AND layers>0 bands), SEPARATE from depth_edge_rel and often the DOMINANT tear cause on deep / narrow-FOV / interior scenes — grazing walls and receding floors trip the default 12x even where continuous, shredding the mesh into 'combs'. Raise to 40-80 to close them; >80 rubber-sheets real foreground silhouettes onto the background. |
| sky_heuristicopt | BOOLEAN | true | layers=0 relief mesh: exclude above-horizon far/rough regions as sky before triangulation. Correct OUTDOORS; turn OFF for INTERIORS (it eats the ceiling / vault / far wall as 'sky', punching large holes). Ignored when sky (the SAM card) is on — that mask governs. (layers>0 bands: sky handled per-band via exclude/scope instead.) |
| normal_edge_degopt | FLOAT | 00–180 | 0 = off. A THIRD tear test on surface-normal BEND (layers=0 relief AND layers>0 bands): tears real creases / occlusion silhouettes while leaving smoothly-receding walls and floors intact (unlike max_edge_factor, which trips on any grazing surface). Pair with a HIGHER max_edge_factor: raise mef to stop comb-tearing continuous grazing surfaces, then set ~40-70 here to keep genuine edges torn. |
| depth_modelopt | COMBO | depth-anything/Depth-Anything-V2-Metric-Outdoor-Large-hf | Monocular depth backend (fed the solved focal). V2-Metric-Outdoor (DEFAULT) = Apache, transformers-only (NO extra install), best all-round on OUTDOOR/sky scenes; V2-Metric-Indoor is its interior twin. MoGe-2 (Ruicheng/moge-*) = MIT, cleanest on ENCLOSED/INTERIOR shots but masks sky (poor outdoors) — needs [moge]. DA3* (EXPERIMENTAL) = strong metric, heavy deps, DA3NESTED is non-commercial CC BY-NC — needs [neural-da3]. Pick per shot: outdoor->V2-Outdoor, interior->MoGe or V2-Indoor. (A/B 2026-07-13.) |
| vlm_scopeopt | BOOLEAN | true | When use_vlm: also SCOPE each band by the VLM's SAM prompt (band ∩ segment). A PARTIAL segment match legitimately keeps the scope and can cut real band geometry — found live on the ghost-town plate, where the mid band's 4.6% segment left the rest of the band exposing the behind-band's fill smear. OFF = VLM still drives bands/geometry, layers stay band-only (robust full coverage — best for camera moves). |
| sky_inpaint_modeopt | COMBO | lama | How to build the sky card's clean plate: lama = fast deterministic edge-fill (needs comfyui-inpaint-nodes), sdxl = generative SDXL inpaint (needs a checkpoint). |
| sky_lama_grow_pxopt | INT | 320–128 | Mask dilation before LaMa sky inpaint. Larger = more aggressive removal of foreground silhouettes from the sky plate. |
| sky_sdxl_checkpointopt | STRING | SDXL/sd_xl_base_1.0.safetensors | SDXL checkpoint filename in models/checkpoints when sky_inpaint_mode=sdxl. |
| sky_sdxl_positiveopt | STRING | clear seamless sky, high detail, no buildings, no trees, no roofs | Positive prompt for SDXL sky inpaint. |
| sky_sdxl_negativeopt | STRING | building, tree, roof, person, vehicle, text, watermark, blurry | Negative prompt for SDXL sky inpaint. |
| sky_sdxl_seedopt | INT | 00–18446744073709550000 | Seed for SDXL sky inpaint. 0 = deterministic default behavior of the sampler node. |
| retopo_methodopt | COMBO | off | Retopologize every relief mesh (single mesh or all bands) via ONE AtlasRetopologizeLayer 🔷 (layer='*') at the end of the chain — the sky dome is never touched. off (DEFAULT) = free; quad/decimate reduce to the vertex budget; voxel_remesh closes interior tears watertight. For `smooth` (interior Taubin) use the standalone 🔷 node — here boundary_smooth_iterations covers the silhouette-rounding case. |
| retopo_target_vertex_countopt | INT | 2000100–200000 | Per-mesh vertex budget for quad/decimate (ignored by voxel_remesh). |
| boundary_smooth_iterationsopt | INT | 00–50 | Taubin-relax every open boundary loop AFTER retopo — rounds lattice-staircase silhouette jaggies; projective UVs regenerated per layer camera. Works with retopo_method=off ('just round the silhouette'). 0 = off. Pair with sub_quad_boundary: smoothing ROUNDS a staircase, the cut RELOCATES it (measured 5.67 -> 1.43px alone, 1.35px together). |
| sub_quad_boundaryopt | BOOLEAN | false | layers=0 relief mesh: cut a torn cell AT the depth cliff instead of deleting the whole cell. Tearing is per grid cell, so a silhouette turns only in whole-cell steps AND a cell of real surface is lost 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 out of 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 - the tear is untouched. Ignored when layers>0 (bands build their own meshes). |
| raw_metaopt | ATLAS_RAW_META | — |
Outputs (5)
| Name | Type | Description |
|---|---|---|
| solve | ATLAS_SOLVE | — |
| image | IMAGE | — |
| depth | ATLAS_DEPTH_MAP | — |
| sky_mask | MASK | — |
| report | STRING | — |