ComfyUI-PanoTools - Panorama Generation Tools

Professional panorama generation tools for ComfyUI. Converts perspective images/video to equirectangular 360° panoramas with automatic camera calibration. Supports multiple output resolutions (3840×1920, 4096×2048, 5760×2880, 7680×3840) and accepts any reasonable aspect ratio (portrait to landscape, 1:2 to 2:1).

Credits: Based on the original perspective-to-panorama method by hybridworkflow.

Note: Includes GeoCalib (ECCV 2024) by ETH Zurich under Apache-2.0 license. See NOTICE and LICENSE-GeoCalib.

Features

• Flexible Input: Accepts any aspect ratio between 1:2 (portrait) and 2:1 (landscape)
• Configurable Output: Choose from 4 resolutions (3840×1920, 4096×2048, 5760×2880, 7680×3840) to balance quality, memory usage, and processing speed
• Automatic Calibration: AI-based (GeoCalib), horizon detection, or vertical line detection. Caution: Horizon and vertical detection mostly untested.
• Manual Control: Fine-tune roll, pitch, and vFoV parameters

Installation

Via ComfyUI Manager (Coming soon):

1. Search for "PanoTools"
2. Install

Via Git:

cd ComfyUI/custom_nodes
git clone https://github.com/9nate-drake/ComfyUI-PanoTools.git

Dependencies:

pip install -r requirements.txt

Required: torch, torchvision, opencv-python, kornia, matplotlib, numpy

GeoCalib: Bundled with the package. Model weights (~110-220 MB) download automatically on first use.

Usage

Node Locations

Camera Calibration: Add Node > PanoTools > Camera Calibration Perspective to Panorama: Add Node > PanoTools > Perspective to Panorama

Camera Calibration Node

Estimates camera parameters (roll, pitch, vFoV) from a single image.

Calibration Methods

| Method | Best For | Technology | |--------|----------|------------| | GeoCalib Auto | Any scene, most accurate | Deep learning (ECCV 2024) | | Horizon Detection | Outdoor scenes with visible horizon | Edge detection + Hough transform | | Vertical Lines | Architectural/urban scenes | Edge detection + Hough transform | | Manual | Known camera parameters | Direct input |

Inputs

• image (IMAGE): Single image or first frame to analyze
• calibration_method: Choose method (geocalib_auto recommended, horizon_detection, vertical_lines, manual)
  • geocalib_auto : AI-based, works for all scenes (recommended)
  • horizon_detection: For outdoor/ocean scenes with visible horizon
  • vertical_lines: For architecture/buildings with vertical structures
  • manual: Specify values manually
• manual_roll/pitch/vfov (optional): Manual override values (only used when method is 'manual')
• horizon_roi_top/bottom (optional): Region of interest for horizon detection
• min_line_length/angle_threshold (optional): Parameters for vertical line detection

Outputs

1. roll (FLOAT): Camera roll in degrees
2. pitch (FLOAT): Camera pitch in degrees
3. vfov (FLOAT): Vertical field of view in degrees
4. focal_length (FLOAT): Focal length in pixels
5. focal_length_mm (FLOAT): 35mm equivalent focal length (e.g., 24mm, 50mm)
6. calibration_info (STRING): Diagnostic information

Perspective to Panorama Node

Converts perspective images/video to 360° equirectangular panoramas.

Inputs

• images (IMAGE): Input frames (any aspect ratio from 1:2 portrait to 2:1 landscape)
• output_resolution (DROPDOWN): Output panorama resolution (default: 5760×2880)
  • 3840×1920: Lower quality, faster processing, less VRAM
  • 4096×2048: Medium-low quality, good for quick previews
  • 5760×2880: Balanced quality/performance (recommended)
  • 7680×3840: Highest quality, requires more VRAM and processing time
• interpolation (DROPDOWN): Interpolation method (default: lanczos)
  • lanczos : Best quality, sharpest edges, reduces jagged artifacts (recommended)
  • cubic: High quality, smooth results, faster than lanczos
  • bilinear: Basic quality, fastest processing
• roll (FLOAT): Camera roll in degrees (tilt left/right, default: 0.53)
• pitch (FLOAT): Camera pitch in degrees (tilt up/down, default: -9.47)
• vfov (FLOAT): Vertical field of view in degrees (default: 52.14)
  • Typical values: 40-60° (standard lens), 70-90° (wide angle), 10-30° (telephoto)
• focal_length (FLOAT): Focal length in pixels (0 = auto-calculate from vfov)

Outputs

1. panorama (IMAGE): Equirectangular panorama at selected resolution
2. mask_single (IMAGE): Valid projection area mask (white=valid, black=invalid)
3. info (STRING): Processing information

Example Workflow

Example workflows available in the /workflows folder.

Automatic Calibration (Recommended):

[Load Video] → [Get Frame 0] → [Camera Calibration (Auto)] → roll/pitch/vfov
                                                                      ↓
[Load Video] ────────────────────→ [Perspective to Panorama] ← ──────┘
                                              ↓
                                        [Save Image]

1. Load video/images
2. Extract one frame for calibration
3. Run Camera Calibration (use geocalib_auto for best results)
4. Connect calibration outputs to Perspective to Panorama node
5. Process all frames

Troubleshooting

Node doesn't appear:

• Restart ComfyUI
• Check for Python errors in console

"ERROR: Input must be between portrait(1:2) and landscape(2:1)":

• Input aspect ratio too extreme (must be 0.5 ≤ width/height ≤ 2.0)

Panorama rotated/off-center:

• Use Camera Calibration node for automatic detection
• Manually adjust roll (horizon level) and pitch (vertical centering)

"ERROR: OpenCV not available" or "ERROR: GeoCalib failed to load":

• Run: pip install -r requirements.txt
• Restart ComfyUI

"WARNING: No horizon/lines detected":

• Try geocalib_auto method instead
• Adjust detection parameters (ROI, min_line_length, angle_threshold)

Jagged edges or aliasing artifacts in panorama:

• Switch from bilinear to cubic or lanczos interpolation
• Lanczos provides the best quality but is slower
• Increase output resolution for better quality at FOV edges

Technical Details

Projection: Reverse equirectangular projection using spherical coordinates and pinhole camera model Coordinate System: X=Right, Y=Up, Z=Forward, ZYX Euler angles Interpolation: OpenCV remap with configurable methods (Lanczos4, Cubic, Bilinear)

• Lanczos4: 8×8 pixel kernel, sharp and artifact-free
• Cubic: 4×4 pixel kernel, smooth and balanced
• Bilinear: 2×2 pixel kernel, fast but basic Optimization: Mask computed once per batch, vectorized NumPy operations, GPU-accelerated interpolation (if OpenCV built with CUDA)

Version History

• v1.0 Initial release

Credits

Original Method: hybridworkflow - Reddit post

GeoCalib (ECCV 2024): ETH Zurich - Paper | GitHub | Demo

Contributing

Contributions welcome! Submit pull requests or open issues for bugs and feature requests.