Extensions/ComfyUI-RMBG
ComfyUI Extension

ComfyUI-RMBG

An all-in-one, ultra-fast ComfyUI custom node for Depth Anything 3 (DA3).

By 1038lab·Created a day ago·Updated a day ago· 0
1038lab/ComfyUI-Depth-Anything-3
Nodes—
On cloudLocal install
Stars0
Updateda day ago
Readme

ComfyUI Depth Anything 3 (DA3)

<p align="center"> <img src="https://img.shields.io/badge/ComfyUI-Custom--Node-blue?style=for-the-badge&logo=comfyui" alt="ComfyUI"> <img src="https://img.shields.io/badge/Depth%20Anything-V3-orange?style=for-the-badge" alt="DA3"> <img src="https://img.shields.io/badge/PyTorch-%3E=2.0-red?style=for-the-badge&logo=pytorch" alt="PyTorch"> <img src="https://img.shields.io/badge/HuggingFace-Models-yellow?style=for-the-badge&logo=huggingface" alt="HuggingFace"> </p>

An all-in-one, ultra-fast ComfyUI custom node for Depth Anything 3 (DA3) by ByteDance Seed. Designed specifically for generating temporally stable, geometrically accurate Depth Video Sequences and single-frame depth maps to guide modern AI Video Generation pipelines (such as Wan 2.1, CogVideoX, HunyuanVideo, Stable Video Diffusion, and ControlNet Depth).

🎯 Why This Node? Supercharging AI Video Generation

Generating coherent video with AI models (e.g. Wan 2.1, CogVideoX, SVD) often suffers from critical spatial issues:

  • Structural morphing: Limbs deform, faces melt, and objects shift unrealistically across frames.
  • Spatial flickering & depth collapse: Backgrounds wobble and fail to maintain consistent perspective when cameras move.
  • Lost edge boundaries: Thin structures like hair, fingers, branches, and poles blend into the background.

The Solution: Accurate Depth Maps as 3D Structural Anchors

By passing video frames through this node, you obtain a temporally consistent, high-contrast Depth Video Sequence:

  1. Precise 3D Geometry: Locks down object bounds, character silhouettes, and true depth planes.
  2. Rock-Solid Camera Motion: Provides rigid geometric conditioning for Depth-to-Video models, eliminating background warping.
  3. Streamlined One-Node Experience: No need to chain separate loaders, preprocessors, and renderers. Plug your video in, select a model, and get your depth stream ready for generation!

📊 Model Comparison: Why Depth Anything 3?

Depth Anything 3 marks a fundamental paradigm shift from previous disparity-based depth models:

| Feature | Depth Anything V1 / V2 | Video-Depth-Anything | Depth Anything 3 (DA3) | | :--- | :---: | :---: | :---: | | Depth Representation | Disparity (Inverse Depth) | Disparity + Temporal Conv | Unified Depth-Ray (True Metric & Relative) | | Physical Geometry | Prone to non-linear distortion | Moderate distortion | True Euclidean depth with natural perspective | | Edge & Thin Structures | Blurry around hair & fingers | Moderate blurring | Razor-sharp edge preservation | | Model Footprint | Large (>300M) | Heavy (Multi-component) | Small (80M) / Base (120M) ultra-lightweight | | Auto-Download | Manual download needed | Manual download needed | 100% Automatic Hugging Face integration | | ComfyUI Integration | Multi-node pipeline | Multi-node pipeline | Single all-in-one node (IMAGE in, IMAGE out) |

🏆 Model Zoo & Selection Guide

All models are automatically downloaded on first run from 1038lab/Depth-Anything-3 directly into ComfyUI/models/geometry_estimation/.

| Model (model) | Parameters | VRAM Footprint | Inference Speed | Recommended Application | | :--- | :---: | :---: | :---: | :--- | | depth_anything_3_small.safetensors | 0.08B | ~2.8 GB | ⚡ Ultra Fast | Ideal for long video batches, real-time preview, and low-VRAM GPUs. | | depth_anything_3_base.safetensors | 0.12B | ~3.6 GB | 🚀 Fast | Best everyday balance of speed, low VRAM, and geometric fidelity. | | depth_anything_3_mono_large.safetensors | 0.35B | ~5.5 GB | 🌟 Standard | Top Recommendation for Video Generation Guiding: Maximum detail on silhouettes, complex textures, and facial features. | | depth_anything_3_metric_large.safetensors | 0.35B | ~5.5 GB | 📐 Standard | True Real-World Metric Units (meters): Essential for 3D reconstruction, VFX compositing, and camera tracking. |

⚙️ Node Parameters Explained (Say Goodbye to Complex Jargon!)

The node is built with simple, intuitive terms and built-in interactive tooltips for every parameter:

| Parameter | Type | Default | Options | What it does (Plain English) | | :--- | :---: | :---: | :---: | :--- | | images | IMAGE | Required | Image / Video frames | The input video frame batch or single image. Connect directly from Load Video (e.g. VHS Video Helper Suite) or Load Image. | | model | COMBO | small | Small / Base / Mono-Large / Metric-Large | Which DA3 model to run. Auto-downloads from Hugging Face if not found on your machine. | | resolution | INT | 504 | 140 ~ 2520 (step 14) | Internal processing size (longest side). 504 is fast and lightweight. Increase to 756 or 1008 for finer edge details. | | normalization | COMBO | min_max | min_max, v2_style, raw | Depth normalization mode:<br>• min_max: Standard 0–1 depth (White=Near, Black=Far). Always use this for AI Video Generation & ControlNet!<br>• v2_style: Enhanced perceptual contrast with sky suppression.<br>• raw: Preserves physical unscaled depth numbers. | | colormap | COMBO | gray | gray, inferno, turbo | Visual color style:<br>• gray: Standard grayscale. Required for Video Gen, ControlNet, and Depth-to-Video.<br>• inferno: Orange-purple heatmap for human inspection.<br>• turbo: Rainbow spectrum preview. | | unload_model | BOOLEAN | False | True / False | Memory management: Unloads the DA3 model and cleans VRAM caches immediately after execution. Highly recommended when chaining into heavy video generation models (Wan 2.1, CogVideoX, HunyuanVideo). | | weight_dtype | COMBO | default | default, fp16, bf16, fp32 | Calculation precision:<br>• fp16: Recommended for Nvidia GPUs (cuts VRAM in half and runs much faster).<br>• bf16: Best for RTX 30xx/40xx series.<br>• fp32: Full precision. |

Output:

  • IMAGE: Standard ComfyUI image batch containing depth frames matching the exact frame count, aspect ratio, and timing of your input. Connect directly to Video Combine, Preview Image, or your video generation model.

⚡ Quick-Start Settings for Beginners

For the best results with Video-to-Video or Depth-to-Video generation:

  • model: depth_anything_3_mono_large.safetensors (for maximum detail) or depth_anything_3_small.safetensors (for fast iterations)
  • resolution: 504 (or 756 for 1080p source)
  • normalization: min_max
  • colormap: gray
  • weight_dtype: fp16

📦 Installation

Clone this repository into your ComfyUI custom_nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/1038lab/ComfyUI-Depth-Anything-3.git

If using a portable or custom Python environment, install required dependencies:

pip install -r requirements.txt

(Standard ComfyUI portable packages already include all required libraries)

Restart ComfyUI and search for Depth Anything 3 in your node menu.

📁 Workflows

Ready-to-use ComfyUI workflow JSON files are available in the workflows/ directory:

  • depth_video_basic.json: Video loader ➔ Depth Anything 3 ➔ Video Combine.
  • depth_guided_video_generation.json: Depth Anything 3 combined with video diffusion models.

If this custom node helps you or you like my work, please give me ⭐ on this repo! It's a great encouragement for my efforts!

📄 License & Acknowledgments

  • Depth Anything 3 architecture & weights developed by ByteDance Seed.
  • Safetensors model checkpoints hosted by 1038lab.
  • Model Licensed under the Apache 2.0 License.
  • GPL-3.0 License