◎ Radiance Depth Map

Depth Anything V2 is the model that ended the "which depth estimator" argument - its Large checkpoint has been the default ControlNet depth preprocessor since 2024, and it's still the daily driver. Radiance Depth Map Generator is just a tidy wrapper around it that runs the model for you, with two useful additions: a video-safe normalization mode, and a one-wire connection into the pack's own Depth of Field node so you can turn a flat image into a defocus shot without leaving the graph.

Feed it an image, it estimates monocular depth - brightness encodes distance, near = white, far = black - and hands you a 3-channel grayscale depth map. That map is the standard ingredient for three things: a ControlNet depth condition, a depth-of-field blur pass, and fake parallax/3D warp.

What you're setting

• model_size - Small (25M), Base (98M), Large (335M). Small for fast previews, Large for best quality. The model downloads on first use from HuggingFace (depth-anything/Depth-Anything-V2-*-hf), and this is the big install gotcha: the first run stalls on a download that can be a few hundred MB.
• normalize - on by default. For single images it maps depth to 0–1. For video it does the smarter thing: frames are standardized per-batch so the depth range doesn't wobble between frames - which is what keeps a depth-driven effect from flickering. That temporal consistency is why this node is worth having over raw controlnet_aux.
• invert - flips the convention (white = far). You'll need this when your downstream node expects the opposite polarity.
• blur_edges - 0 to 5, a Gaussian blur over the depth map to smooth discontinuities. This is the fix for the classic depth-artifact where a hard edge in the map creates a halo in your DoF pass.
• use_gpu - on by default; drop it for CPU-only rigs.

Output is a single depth_map IMAGE, ready for a ControlNet apply or straight into Radiance Depth of Field.

The license trap

Read this before you build a product on it: the Large and Base V2 weights are non-commercial (the V2 Small and V1 are Apache 2.0). The Large is the one the node recommends, and it's exactly the one you can't legally ship in a paid tool. If that matters, you're either on Small or you're swapping in a commercial-friendly model and using this node only as a preview.

Honest notes

The pack's launch thread was met with real skepticism about the 32-bit claims, and a depth map is the perfect example of a node that doesn't need them - it outputs a normalized grayscale image, so the float pipeline is mostly irrelevant here. It works as advertised. The bigger practical caveats are the first-run model download and the V2 non-commercial license, both of which are easy to discover the hard way.

Install

In the fxtdstudios/radiance pack - Manager search "Radiance", or:

cd ComfyUI/custom_nodes
git clone https://github.com/fxtdstudios/radiance.git
cd radiance
pip install -r requirements_linux.txt   # or windows/mac_silicon

Restart after; Linux needs libopenexr-dev before pip. The pack pulls transformers (the depth model runs through it), so the dependency install is heavy and the first boot is slow.