Nodes/ComfyUI-Depth-Anything-3/Depth Anything 3
ComfyUI Node

Depth Anything 3

A ComfyUI node in šŸ§ŖAILab/Geometry with 7 inputs and 1 output.

By 1038labĀ·Created 4 days agoĀ·Updated 3 days agoĀ· 1
Depth Anything 3
  • images
  • IMAGE
ā—„modeldepth_anything_3_small.safetensorsā–ŗ
ā—„resolution512ā–ŗ
ā—„normalizationmin_maxā–ŗ
ā—„colormapgrayā–ŗ
ā—„unload_modelfalseā–ŗ
ā—„weight_dtypedefaultā–ŗ
CategoryšŸ§ŖAILab/Geometry

Inputs (7)

NameTypeDefaultDescription
imagesIMAGEInput video frames or single image. Connect from Load Video (e.g. VHS Video Helper Suite) or Load Image.
modelCOMBOdepth_anything_3_small.safetensorsDepth Anything 3 model weight (auto-downloads if missing): - small: Fastest, lowest VRAM (~0.08B params, great for long videos & quick testing) - base: Great balance between speed and precision (~0.12B params) - mono_large: Highest visual depth detail (~0.35B params, best choice for AI video generation guiding) - metric_large: Real-world metric depth in meters (~0.35B params)
resolutionINT512128ā€“2560Model internal processing resolution (longest side, multiple of 14): - 504: Recommended default (fast, low VRAM) - 756 / 1008: Sharper edges & finer depth details (requires more VRAM) The output is always automatically upsampled back to the original input resolution.
normalizationCOMBOmin_maxDepth range normalization mode: - min_max: Standard 0 to 1 range (near=white 1.0, far=black 0.0). Required for AI video generation (Wan/CogVideoX/Hunyuan/SVD) and ControlNet! - v2_style: Adaptive contrast balance with automatic sky clipping - raw: Unscaled numerical depth values (preserves metric units for 3D/VFX workflows)
colormapCOMBOgrayOutput color format: - gray: Standard grayscale (Required for AI Video Generation, ControlNet, and Depth-to-Video) - inferno: High-contrast orange/purple thermal heatmap (best for visual human inspection) - turbo: Smooth rainbow colormap (visual human inspection)
unload_modelBOOLEANfalseUnload model from VRAM immediately after execution to free maximum GPU memory for downstream models (e.g. Wan 2.1, CogVideoX, HunyuanVideo).
weight_dtypeCOMBOdefaultComputation precision: - default: Uses model native weights - fp16: Recommended for Nvidia GPUs (saves ~50% VRAM, runs faster) - bf16: Recommended for RTX 30xx/40xx & newer GPUs - fp32: Full precision (uses highest VRAM)

Outputs (1)

NameTypeDescription
IMAGEIMAGEā€”