Multipass Estimate

This is the ambition node of the pack. Give it a plate - a photo, a render, a generated frame - and it produces depth, camera-space position, normals, AO, curvature, albedo, roughness, metallic, and separate diffuse and specular lighting. That's most of what a compositor would ask a renderer for, estimated from one image.

It's also the node with the largest downloads and the most honest "results will vary" warning attached to it. Both of those are true, and they don't cancel out.

What's actually running

Two real models, and they're the reason to take this seriously:

• MoGe-2 (Microsoft Research, with Tsinghua) for geometry - depth, position, normals, AO, curvature. MoGe is the monocular geometry model that predicts an affine-invariant point map rather than a depth map: a per-pixel 3D coordinate, plus a validity mask that excludes sky and other regions where geometry is undefined. From that point map you get metric-scale depth, camera-space position, and normals that are derived from geometry rather than hallucinated texture. MoGe-2 adds metric scale and sharper detail. 662 MB, and the pack says MoGe-2 specifically is required for normals.
• Marigold IID (ETH Zurich) for materials and lighting - albedo, roughness, metallic from the Appearance model (2.5 GB), diffuse and specular light from the Lighting model (1.7 GB, sharing files with Appearance). Marigold is the Stable-Diffusion-derived estimator: instead of training a discriminative model, it reframes the task as a denoising diffusion process on top of SD's visual prior, which is exactly why it generalises to illustrations and synthetic plates that would break a photo-trained model. The cost is speed, and here the cost is real.

The author's framing - "No image-filter guesses" - is the point. This isn't four blurs and a threshold dressed up as passes.

Inputs and outputs

Required: beauty (the plate, one image or a batch), beauty_encoding (sRGB (display) for ordinary images and video frames, Linear (scene) for scene-linear EXR plates - values above 1.0 get clipped for the models and excluded from the lighting fit), and the three toggles: geometry, materials, lighting. Turn off what you don't need; that's the difference between a 662 MB and a ~4 GB download and a two-minute versus a ten-minute run.

The optional controls that matter: geometry_detail (0 fastest to 9 most detail - this is MoGe's resolution level and the main quality/speed dial for the geometry half), fov_x_degrees (set a known field of view, or leave 0 to estimate), normal_convention (OpenGL vs DirectX - flip green to match whatever reads the pass downstream), ao_radius_m and ao_quality for ambient occlusion, material_steps (4 is Marigold's trained default - don't crank it expecting free quality), material_resolution (768 is the trained size), and ensemble_size (average several predictions; slower, steadier).

seed deserves a sentence: it's reused on every frame deliberately, to limit flicker. And batch_is_sequence (default true) treats a batch as consecutive frames so motion vectors make sense. For a single still, prev_frame gives the motion vector something to compare against.

Outputs: passes (a RADIANCE_PASSES bundle for Write EXR Passes and Relight) plus thirteen individual IMAGE outputs - beauty (linear), albedo, roughness, metallic, diffuse_lighting, specular_lighting, normal_map, depth, world_position, ao, curvature, motion_vector, geometry_mask - and info, a JSON string with the models used, estimated field of view, lighting scale factors and fit error. Read info. The fit error tells you whether the lighting decomposition worked or just found a plausible-looking split, and the FOV tells you whether your metric depth is trustworthy.

Installing Radiance, and the model download

Manager → search Radiance → install → restart → refresh. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/fxtd-studios/radiance.git
cd radiance
python -m pip install -r requirements.txt

Windows portable: use python_embeded\python.exe. Then the important part: this node needs diffusers and accelerate (the latter so Marigold's 1.7 GB UNets fit without a second full-size copy in RAM), and it downloads weights on first run. download_missing_models defaults on and pulls pinned versions from Hugging Face. RADIANCE_ALLOW_DOWNLOADS=0 or HF_HUB_OFFLINE=1 always blocks it, which is what you want on a metered connection. You can also bring your own MoGe via the moge_model input from ComfyUI's Load MoGe Model - but it must be MoGe-2 for normals.

Troubleshooting

Video flickers. Documented, expected. These are per-frame estimates; the fixed seed limits it and ensemble_size steadies it, at the cost of time.

Sky and edges of frame look wrong. Check geometry_mask - sky is where MoGe found no surface, and depth is 0 there by design.

It ran out of memory or took forever. Marigold is expensive. Drop material_resolution toward 768 (or below), keep material_steps at 4, and consider turning lighting or materials off. On consumer cards the geometry half is the affordable one.

Expectations, from people who actually ran it. A user in the community discussion of this pack noted the practical ceiling: "With a 4090 you won't be able to run the giant models in fp16 when it goes above 1-2k res," and that these nodes "won't create real 16/32 bit exr. For that you need your input to have the data." That's the right way to read this node. It estimates passes - usefully, from real models - but it doesn't recover information that was never captured. If you have genuine multilayer EXRs, use Multipass AOV Reader instead and skip the guesswork entirely.

Licensing, briefly. Radiance code is GPL-3.0. Check the licence on each model you enable; MoGe is permissively licensed, and Marigold is Apache-2.0, but verify before shipping anything commercial.