ComfyUI Node
Atlas Depth Map 🌊
A ComfyUI node in Atlas/03 · Depth with 10 inputs and 1 output.
Atlas Depth Map 🌊
- image
- solve
- depth
◄depth_modeldepth-anything/Depth-Anything-V2-Metric-Outdoor-Large-hf►
◄deviceauto►
◄moge_resolution_level9►
◄moge_max_side0►
◄moge_checkpoint_path►
◄moge_tile_side0►
◄moge_tile_overlap0.25►
◄moge_report_free_focalfalse►
CategoryAtlas/03 · Depth
Inputs (10)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | — | |
| depth_modelopt | COMBO | depth-anything/Depth-Anything-V2-Metric-Outdoor-Large-hf | 10 options: depth-anything/Depth-Anything-V2-Metric-Outdoor-Large-hf, depth-anything/Depth-Anything-V2-Metric-Indoor-Large-hf, depth-anything/DA3METRIC-LARGE, depth-anything/DA3MONO-LARGE, depth-anything/DA3NESTED-GIANT-LARGE-1.1, Ruicheng/moge-2-vitl-normal, +4 |
| deviceopt | COMBO | auto | 4 options: auto, cuda, mps, cpu |
| solveopt | ATLAS_SOLVE | Optional — supplies the SOLVED focal (GeoCalib/VP) for DA3METRIC's canonical→metric conversion (focal_source='solve' instead of the assumed normal-lens fallback). Ignored by V2 models. | |
| moge_resolution_levelopt | INT | 90–9 | MoGe ONLY. Its own token-budget dial; 9 is MoGe's default and full detail. Lower = faster and coarser. No effect on DA/V2/DepthPro. |
| moge_max_sideopt | INT | 00–16384 | MoGe ONLY. Cap the longer edge BEFORE inference (0 = off). MoGe resamples internally anyway, so this buys VRAM and time, not quality — an 8K plate is ~415 MB of float32 on the device before MoGe touches it. Depth/normals still come back at SOURCE size. |
| moge_checkpoint_pathopt | STRING | MoGe ONLY. Local MoGe `model.pt` to load instead of downloading from HuggingFace (air-gapped / shared model dirs). NOT ComfyUI core's geometry_estimation/*.safetensors — different container, and it carries no model_config. | |
| moge_tile_sideopt | INT | 00–4096 | MoGe ONLY. Run inference on overlapping TILES of this size at SOURCE resolution (0 = off). The opposite lever to moge_max_side: that one downscales to save VRAM, this one refuses to downscale so a 36MP plate keeps its fine structure — the model spends its whole token budget on each tile instead of on the shrunken whole. Costs one inference pass PER TILE plus one global pass, so a 4x4 tiling is ~17x the time. Every tile is affine-fitted onto that global pass first: monocular depth is scale-ambiguous per input, so raw tiles disagree and pasting them steps at every seam. Try 1024. |
| moge_tile_overlapopt | FLOAT | 0.250–0.5 | MoGe ONLY. Tile overlap as a fraction of tile size. More overlap = wider blend and more tiles (slower). Only used when moge_tile_side > 0. |
| moge_report_free_focalopt | BOOLEAN | false | MoGe ONLY. When a solve is wired, MoGe is FED the solve's focal, so its own intrinsics just echo it. This runs a second, fov-free pass (depth discarded, cheaper resolution level) and records MoGe's INDEPENDENT focal so the 🩺 debug report can flag a focal_mismatch (band 0.75-1.33). ~+30% time. sh001 example: solve 6207 px vs MoGe 5278 px — invisible without this. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| depth | ATLAS_DEPTH_MAP | — |