Extensions/Dobidop ComfyStereo
ComfyUI Extension

Dobidop ComfyStereo

Two simple nodes for stereoscopic image generation. Nodes: Stereo Image Node - a basic port from the Automatic1111 stereo script in…

By Dobidop·Created 2 years ago·Updated 3 months ago· 52
Dobidop/ComfyStereo
Nodes5
On cloudLocal install
Categorystereo/native, image/stereo
Stars52
Updated3 months ago

Nodes (5)

Native Stereo Image Viewer (PyOpenXR)
stereo/native
Native Stereo Video Viewer (PyOpenXR)
stereo/native
Check Native VR Status
stereo/native
StereoDiffusion
image/stereo
Stereo Image Node
sd
Readme

ComfyStereo - Stereoscopic 3D Toolkit for ComfyUI

A stereoscopic 3D toolkit for ComfyUI that combines three solutions into one unified package:

  1. Stereo Image Generation - Depth-based stereo conversion with GPU acceleration
  2. Native VR Viewing - PyOpenXR viewer for direct VR headset viewing
  3. StereoDiffusion - AI-powered stereo generation using diffusion models

Example Workflow

This example demonstrates the full stereo conversion pipeline using a depth map and video input.

Example Workflow Graph

<img src="examples/image-workflow.png" alt="drawing" width="500"/>

Download workflow:
Video2Stereo.json — Import directly into ComfyUI

Final Stereo Output

This is the final stereoscopic video generated by ComfyStereo.

Stereo Video

Direct link:
stereovideo.mp4

Original Input Video

This is the original monoscopic input video before stereo processing.

<img src="examples/example-video.gif" alt="drawing" width="200"/>

Direct link:
example-video.mp4

Depth Map Used

This is the depth map generated and used in the workflow.

<img src="examples/depthmap_video.gif" alt="drawing" width="200"/>

Direct link:
depthmap_video.webm

Features Overview

Core Stereo Generation

  • GPU-Accelerated Processing - 5-20x faster depth processing with CUDA
  • Advanced Fill Techniques - Multiple interpolation methods (Polylines, Naive, Hybrid Edge, GPU Warp)
  • Edge-Aware Depth Blurring - Reduces artifacts at high divergence settings
  • Multiple Output Formats - Side-by-Side, Top-Bottom, Red-Cyan Anaglyph
  • Batch Video Processing - Memory-efficient video frame processing

Native VR Viewing

  • Auto-Launch to Headset - Direct VR viewing without browser
  • Multiple Stereo Formats - Side-by-Side, Over-Under, Mono
  • Projection Options - Flat, Curved, 180° Dome, 360° Sphere
  • Image & Video Support - View both stereo images and videos
  • All VR Headsets - Quest, Vive, Index, WMR, and more

StereoDiffusion AI

  • AI-Powered Generation - Uses diffusion models for stereo creation
  • DDIM Inversion - Null-text optimization for high-quality reconstruction
  • Bilateral Neighbor Attention - Stereo-consistent diffusion
  • ComfyUI Native Models - Works with MODEL/CLIP/VAE inputs
  • Diffusers Support - Also works with HuggingFace model IDs

Installation

Method 1: ComfyUI Manager (Recommended)

  1. Open ComfyUI Manager
  2. Search for "ComfyStereo"
  3. Click Install
  4. Restart ComfyUI

Method 2: Manual Installation

  1. Clone the repository:
cd ComfyUI/custom_nodes/
git clone https://github.com/Dobidop/ComfyStereo.git
cd ComfyStereo
  1. Install dependencies (choose one):
# Base only (stereo generation + DeoVR)
pip install -r requirements.txt
  1. Restart ComfyUI

Available Nodes

Stereo Image Generation Nodes

1. Stereo Image Node

The main node for depth-based stereo conversion.

Inputs:

  • image (IMAGE) - Source image
  • depth_map (IMAGE) - Depth map (grayscale)
  • divergence (FLOAT) - Stereo effect strength (0.05-15.0, default: 3.5)
  • separation (FLOAT) - Additional horizontal shift (-5.0 to 5.0)
  • stereo_balance (FLOAT) - Effect distribution between eyes (-0.95 to 0.95)
  • convergence_point (FLOAT) - Depth level at screen plane (0.0-1.0, default: 0.5)
  • modes - Output format: left-right, right-left, top-bottom, bottom-top, red-cyan-anaglyph
  • fill_technique - Infill method (see Infill Methods)
  • depth_map_blur (BOOLEAN) - Enable edge-aware depth blurring
  • depth_blur_edge_threshold (FLOAT) - Gradient sharpness cutoff (0.1-15.0)
  • batch_size (INT) - Frames per memory cleanup cycle

Outputs:

  • stereoscope (IMAGE) - Final stereo image
  • blurred_depthmap_left (IMAGE) - Processed left depth map
  • blurred_depthmap_right (IMAGE) - Processed right depth map
  • no_fill_imperfect_mask (MASK) - Unfilled region mask

Native VR Viewer Nodes

2. Native Stereo Image Viewer

Auto-launches images directly into VR headset.

Inputs:

  • image (IMAGE) - Stereo image
  • stereo_format - Side-by-Side, Over-Under, Mono
  • projection_type - Flat Screen, Curved Screen, 180° Dome, 360° Sphere
  • screen_size (FLOAT) - Virtual screen size (1.0-10.0)
  • screen_distance (FLOAT) - Distance from viewer (1.0-10.0)
  • swap_eyes (BOOLEAN) - Swap left/right
  • auto_launch (BOOLEAN) - Launch into headset
  • background_color - Black, Dark Gray, Gray, White

Outputs:

  • passthrough (IMAGE) - Original image

3. Native Stereo Video Viewer

Play stereo videos in VR with keyboard controls.

Inputs:

  • video_path (STRING) - Path to stereo video
  • stereo_format - Side-by-Side, Over-Under, Mono
  • projection_type - Flat Screen, Curved Screen, 180° Dome, 360° Sphere
  • screen_size (FLOAT) - Virtual screen size
  • screen_distance (FLOAT) - Distance from viewer
  • swap_eyes (BOOLEAN)
  • loop_video (BOOLEAN)
  • auto_launch (BOOLEAN)
  • background_color

4. Native VR Status

Check PyOpenXR installation and VR runtime availability.

Outputs:

  • status_message (STRING) - Diagnostic information
  • is_available (BOOLEAN) - VR readiness

StereoDiffusion AI Nodes

5. StereoDiffusion Node

AI-powered stereo generation using diffusion models.

Inputs:

  • image (IMAGE) - Source image
  • depth_map (IMAGE) - Depth map
  • scale_factor (FLOAT) - Disparity strength (1.0-20.0, default: 9.0)
  • direction - "uni" (unidirectional) or "bi" (bidirectional) attention
  • deblur (BOOLEAN) - Add noise to unfilled regions
  • num_ddim_steps (INT) - DDIM steps (10-100, default: 50)
  • null_text_optimization (BOOLEAN) - Enable for better quality (slower)
  • guidance_scale (FLOAT) - CFG scale (1.0-20.0, default: 7.5)
  • model_id (STRING) - HuggingFace model ID (fallback if MODEL/CLIP/VAE not provided)

Outputs:

  • stereo_pair (IMAGE) - Side-by-side stereo image
  • left_image (IMAGE) - Left eye view
  • right_image (IMAGE) - Right eye view

Supports: SD1.x and SD2.x models (SDXL/FLUX planned)

Key Parameters Explained

Divergence

Controls the strength of the 3D effect. Higher values = more depth perception.

  • Low (1-3): Subtle depth
  • Medium (3-7): Balanced effect
  • High (7-15): Extreme pop-out

Convergence Point

Controls which depth appears at screen plane (zero parallax).

  • 0.0 = Nearest depth at screen → Content recedes behind screen
  • 0.5 = Mid-depth at screen → Balanced (default)
  • 1.0 = Furthest depth at screen → Content pops toward viewer

Use cases:

  • Pop-out mode (1.0): Product displays, comics
  • Window mode (0.0): Subtle depth, natural recession
  • Portrait mode (0.6-0.7): Face at screen, background recedes
  • Landscape mode (0.3-0.4): Foreground pops, horizon recedes

Stereo Balance

Distributes divergence between eyes.

  • 0.0 = Even distribution
  • Positive/negative = Shift effect toward one eye

Separation

Additional horizontal shift percentage (independent of depth).

GPU Acceleration

Depth processing is automatically GPU-accelerated when CUDA is available:

  • 5-20x faster blur operations
  • Automatic fallback to CPU if GPU unavailable
  • Zero configuration - works out of the box

Native VR Setup

Requirements

  1. Install PyOpenXR dependencies:
pip install -r requirements.txt

  1. Install a VR runtime:

    • SteamVR (recommended) - Supports most headsets
    • Oculus Runtime - For Meta Quest headsets
    • Windows Mixed Reality - Built into Windows 10/11

  2. Connect your VR headset

Supported Headsets

  • Meta Quest (1, 2, 3, Pro)
  • HTC Vive / Vive Pro
  • Valve Index
  • Windows Mixed Reality headsets
  • Any OpenXR-compatible device

StereoDiffusion Setup

Requirements

  • CUDA-capable GPU with 8GB+ VRAM (16GB recommended)
  • Python 3.8+
  • PyTorch 2.0+

First Run

  • Downloads Stable Diffusion model (releaseversion SD1.5) (~4GB) if not cached
  • Null-text optimization takes ~2-3 minutes on modern GPU
  • Model is cached for faster subsequent runs

Performance Tips

  • Lower num_ddim_steps to 30 for faster processing
  • Disable null_text_optimization for 3x speed (lower quality)
  • Use guidance_scale 3-5 to reduce "burned" look

Troubleshooting StereoDiffusion

  • Out of Memory: Reduce num_ddim_steps, close other apps
  • Black Output: Check depth map is valid grayscale
  • Poor Quality: Enable null_text_optimization, increase num_ddim_steps

License

MIT License - see LICENSE file for details.

Note: This project includes code from multiple sources:

Credits

Created by Dobidop

Acknowledgments

StereoDiffusion

@inproceedings{wang2024stereodiffusion,
  title={StereoDiffusion: Training-Free Stereo Image Generation Using Latent Diffusion Models},
  author={Wang, Lezhong and Frisvad, Jeppe Revall and Jensen, Mark Bo and Bigdeli, Siavash Arjomand},
  booktitle={CVPR},
  year={2024}
}

Prompt-to-Prompt

@article{hertz2022prompt,
  title={Prompt-to-prompt image editing with cross attention control},
  author={Hertz, Amir and Mokady, Ron and Tenenbaum, Jay and Aberman, Kfir and Pritch, Yael and Cohen-Or, Daniel},
  year={2022}
}

Contributing

Contributions welcome! Please submit issues or pull requests.

Support