Nodes/comfyui-timesaver/TS Image Depth
ComfyUI Node

TS Image Depth

Depth map for a still, or for a batch of pictures that have nothing to do with each other. Runs Depth Anything V2 on every picture on its own, the same way the reference implementation does. For a video, where the map has to stay steady from frame to frame, use TS Video Depth instead.

By AlexYez·Created 2 years ago·Updated 3 days ago· 11
TS Image Depth
  • images
  • image
model_filenamedepth_anything_v2_vitl_fp16.safetensors
max_res-1
precisionfp16
colormapgray
CategoryTS/Image/Depth

Inputs (5)

NameTypeDefaultDescription
imagesIMAGEPicture, or a batch of pictures, as IMAGE (N, H, W, 3) in 0..1. Every picture is processed and normalized on its own — a batch here means unrelated images, not a video.
model_filenameCOMBOdepth_anything_v2_vitl_fp16.safetensorsDepth Anything V2 checkpoint, downloaded on first use to ComfyUI/models/depthanything. fp16 safetensors is half the download and loads far quicker; measured against pure fp32 it differs by 0.02% of the depth range, which is nothing.
max_resINT-1-1–8192Longest side the picture is processed at, snapped down to a multiple of 14 (the DINOv2 patch grid). The only quality control here. • -1 (default) — native resolution, exactly what the reference implementation does. Nothing is resampled, so nothing is lost. • 1540 — noticeably quicker on large photos and still sharp; the map comes back up to full size bilinearly. • lower — when VRAM is tight. Depth Anything V2 stops gaining much above roughly 2000 px, so raising this without limit buys little. On out-of-memory the node retries at half the size, and again, logging each step.
precisionCOMBOfp16Inference dtype. • fp16 (default) — 2x faster, ~50% less VRAM. • fp32 — measured difference against fp16 is 0.02% of the depth range. Worth it only if you are chasing the last bit on a smooth surface and have the VRAM to spare.
colormapCOMBOgrayOutput color mapping. • gray (default) — the raw normalized depth in all three channels, which is what the reference returns. Use this when the map feeds another node (ControlNet, 3D, and so on). • inferno / viridis / plasma / magma / cividis — perceptually uniform matplotlib colormaps, for looking at. Bilinear LUT interpolation removes 8-bit banding.

Outputs (1)

NameTypeDescription
imageIMAGEDepth map as IMAGE (N, H, W, 3) float in 0..1 at the original resolution. An ordinary IMAGE — feed it straight into ControlNet, a save node, and so on.