ComfyUI Node RenderPasses Released: Depth, Normals and More From a Single Background Image, Output to Nuke EXR

What happened

"ComfyUI-RenderPasses," released on GitHub by neuropixelplus, is a ComfyUI node that converts anime and VFX background images (plates) into render passes usable for compositing. The input is a flat image with no depth or normals, such as a painting, a generated image or an approved background. According to the project, it reconstructs passes from these and packages them into a single multilayer EXR and a Nuke script that is ready to use as soon as it is opened.

Processing handles the backgrounds in a folder together and proceeds in two stages. In the first stage, "Look," a single master prompt is applied to all backgrounds to relight or finish them. The engine can be chosen from Krea 2, Klein 3, Fibo 2, Krea 2 V2, MOSS and others, and 4K upscaling with SeedVR2 can also be used as an option. In the second stage, each pass is generated from the selected image. Outputs are per-background files plus pases_nuke.exr and pases.nk.

If no background exists yet, the "Create" button, which uses Krea 2, can generate a new one from a prompt. The repository's release is v0.1.0 (2026-10-30), which is the first public version. The license is AGPL-3.0, and the star count was 1 at the time of retrieval. The required ComfyUI version is stated as 0.43 or later.

How it works

Each pass is created by a separate model. Depth (16-bit) is estimated with Depth Anything V2, with an edge-aware correction applied. Normals and albedo are created with a Marigold v2 LoRA running on Qwen-Image-Edit. Raw lighting is obtained by dividing the beauty image by the albedo.

MoGe-2 handles position (P) and the camera lens. It estimates a metric point map, field of view and focal length, and these are written to EXR with the bundled RenderPassesMoGeEXR node. Qwen-Image-Layered or Stable-Layers is used for layer decomposition.

For parallax planes, the hidden regions behind each plane sliced by depth are painted in with Klein to make a 2.5D preview. Automatic objects/Cryptomatte (experimental) work as follows: a local VLM lists the elements, and SAM 3 creates the masks. Fog and cavity are computed on the CPU from depth and normals.

For production use, there is a method that creates Cryptomatte from PSD layers. Each visible leaf layer becomes one CryptoObject, and its name includes the group hierarchy. Coverage is obtained by compositing layer alpha × layer mask × opacity from front to back in the PSD's stacking order. Each pixel can hold up to 4 IDs. If the PSD is larger than the camera frame, the frame position is found by fitting SIFT features with scale and shift, and all layers are cropped to that frame.

Instead of a PSD, you can also supply single-color ID mask PNGs and a legend (JSON). AI light passes (clay, fill/GI, occlusion, shadow, specular, emission, and mattes for characters, props and plants) are created with an image editing model and aligned to the original image. For 4K versions of data-type passes (depth, normals, mattes), the project says it uses ordinary resampling so that no new texture is generated.

By the numbers

Why it matters

In anime and VFX production, approved backgrounds are sometimes delivered as a single flat image. In that case, adding relighting, fog, blur or camera movement on the compositing side is difficult. RenderPasses offers a way to create compositing passes from such flat material.

Because the output assumes Nuke scripts and Cryptomatte, the setup is easy to fit into existing compositing workflows. Meanwhile, the developers position the method that uses PSD-derived layers as better suited to production than automatic estimation in terms of edge quality.

It is also a useful reference for pulling depth, position and parallax planes from a single image in game background production and around generative 3D. However, whether it can be used depends heavily on whether the incorporated models permit commercial use.

How to try it

1. Run git clone of ComfyUI-RenderPasses into the custom_nodes folder of ComfyUI (0.43 or later) and install requirements.txt with pip. After restarting ComfyUI, a RenderPasses button appears in the top bar.
2. When you open the app, a setup check runs and lists the missing models and where to put them (folders under models/). You only need to install the models required for the engine and passes you use.
3. Specify the path of the background folder, or drag and drop images. Choose the master prompt, engine and resolution, press Render, select the images you like, and run Extract passes. Finally, download pases_nuke.exr and pases.nk and open them in Nuke.
4. If you need production mattes, specify the corresponding PSD from "Layers from PSD / mask…" for each background. Install psd-tools beforehand. Detailed steps are in docs/GUIDE.md in the repository.

Caveats and open questions

• The project states that model licenses are the user's responsibility. Its table lists the Klein 3 family as non-commercial (FLUX Kommunity) and Fibo 2 as CC BY-NC, so whether commercial use is possible depends on the engine you choose.
• The RenderPasses table lists the Depth Anything V2 license as Apache-2.0. However, the Depth Anything V2 README says Small is Apache-2.0, while Base/Large/Giant are CC-BY-NC-4.0. Check the model's license before use to see which applies to vitl (Large), which the table refers to.
• Automatic Cryptomatte is described as "experimental," and the project says its edges are not production quality. Quantitative evaluations such as accuracy, speed and required GPU memory could not be confirmed in the material retrieved.
• The retrieved text ends partway, so the latter part of the PSD feature and the output specification for 3D Gaussian splats could not be confirmed. The description and release notes mention 3D splats, but the details are unconfirmed.
• This is the first public release, v0.1.0, with a star count of 1. It is reasonable to treat it as roughly beta-level, and operational stability is unconfirmed.
• Beauty HDR assumes an LTX module that must be prepared separately; it is not bundled.

Glossary

Render pass

An image in which the beauty image is split into elements such as depth, normals, albedo and shadows. Each can be adjusted individually during compositing.

Cryptomatte

A mechanism that stores per-object or per-material IDs and coverage in EXR so that any object can be selected in compositing.

Albedo

The original color of an object's surface, without the effect of lighting.

Multilayer EXR

A high-dynamic-range image format that can store multiple passes in one file.

Nuke

A node-based compositing application widely used in video production.

Depth Anything V2

A foundation model that estimates relative depth from a single image. It comes in Small, Base, Large and Giant sizes.

MoGe-2

A model that estimates a metric point map, field of view and focal length from a single image.

Sources

Primary sources (publisher, paper, repository)

• neuropixelplus/ComfyUI-RenderPassesGitHub（github.com）・en
• DepthAnything/Depth-Anything-V2GitHub（github.com）・en

This article was generated in Japanese by AI from the sources listed (primarily primary sources) and published automatically, without prior human review, after automated checks such as source matching (how we use AI). Figures and specifications are as stated by the publishers and have not been independently verified by us. Please consult the original sources. To report an error, use our contact form; we will review it promptly and correct it. This page is an AI translation of the Japanese original. If anything differs, the Japanese version prevails. Japanese original