Render Depth Anything 3

DA3Render is the visualization stage for the Depth Anything 3 family: feed it a da3_geometry packet from DA3Inference and get back a plain IMAGE - a depth map, a confidence map, or a sky mask. It looks like the boring sibling of the family, but it's where you actually see whether the model did its job, and it's the node whose outputs you'd feed into a depth ControlNet or a parallax effect.

What you can render

• depth - normalized greyscale depth, and depth_colored - the same through the Turbo colormap, which is the canonical "depth map" look.
• sky_mask - sky probability in [0, 1] (Mono/Metric models only).
• confidence - normalized depth confidence (Small/Base models only). Handy for spotting where the model is guessing: dark regions on a confidence render are where the mesh will lie to you.

The output dropdown is a dynamic combo, so picking depth or depth_colored reveals two extra controls worth knowing:

• normalization - v2_style (default, mean/std normalization for perceptually balanced results), min_max (stretch the full range for maximum contrast), or raw (no scaling - for Metric models, this preserves depth in actual metres).
• apply_sky_clip (default off) - clips sky-region depth to the 99th percentile of foreground depth before normalizing. Sky is infinitely far away and will otherwise blow out the whole contrast of your map; with Mono/Metric models you should consider turning this on.

Wiring and gotchas

Output is a standard IMAGE, so it drops straight into Save Image, a ControlNet, or any image consumer. Two failure modes are worth knowing because they're loud and obvious. Request sky_mask from a Small/Base run and you get an error - those models have no sky output, and it will tell you to use Mono/Metric. Same story the other way for confidence. It's not a bug, it's the model capability matrix leaking into the UI. And on the depth side, the "correct" look depends on your normalization: min_max gives a dramatic high-contrast map, but v2_style is the better default for feeding downstream tools because it's balanced rather than dominated by the farthest pixel.