SyncLight: Single-Edit Multi-View Relighting

Official implementation of SyncLight: Single-Edit Multi-View Relighting.

Authors: David Serrano-Lozano, Anand Bhattad, Luis Herranz, Jean-François Lalonde, Javier Vazquez-Corral

---

Overview

SyncLight relights a set of multi-view images consistently from a single light edit. Given a reference image, additional viewpoints, and a target lightmap, the model generates all views under the specified lighting using a multi-view diffusion model (Latent Bridge Matching with MVDream-style 3D attention).

Installation

git clone https://github.com/CVC-Color/synclight
cd synclight
pip install -e .

Python 3.10+ and a CUDA-capable GPU are required. For CPU-only inference use --device cpu (slow).

Model Weights

Download the pretrained weights and place them in ckpt/:

ckpt/
  config.yaml        # already included
  synclight.ckpt     # download separately

Download link: [TODO]

Usage

Lightmap Format

A lightmap is a numpy array of shape (H, W, 4) with float32 values:

Channel	Meaning	Range
0	Light activation (0 = off, 1 = on)	0 or 1
1	Lightness L (normalized)	−1 to 1
2	Color a (normalized)	−1 to 1
3	Color b (normalized)	−1 to 1

L, a, b follow the CIE LAB color space with normalization:

• L_norm = L/50 − 1 where L ∈ [0, 100]
• a_norm = a/128 where a ∈ [−128, 127]
• b_norm = b/128

Creating Lightmaps

Use the interactive lightmap editor:

python lightmap_creator.py --image path/to/reference.png

Controls:

• Left-click / drag — place a circular light source
• Scribble mode — click-drag to draw a freeform light
• Right-click — delete a light
• Scroll — resize selected light
• ab picker — click the color picker panel to set hue
• Temperature slider — set warm/cool color temperature
• Save NPY button — write the lightmap to disk

The editor saves a .npy file and a preview .png in the current directory.

Gradio Demo

python gradio_demo.py --model_weights ckpt/ --device cuda --port 7860

Open http://localhost:7860. Load example images from the sidebar or upload your own images and a lightmap.

Command-Line Inference

python test.py \
  --image_paths input_images/example_1_0_ref.png \
                input_images/example_1_1.png \
                input_images/example_1_2.png \
  --lightmap_path input_images/example_1_0_lightmap.npy \
  --model_weights ckpt/ \
  --output_dir outputs/ \
  --num_inference_steps 1 \
  --device cuda

Output images are saved to outputs/output_0.jpg, output_1.jpg, etc. — one per input view.

Options:

Flag	Default	Description
--image_paths	required	Input images (reference first, then additional views)
--lightmap_path	required	Path to .npy lightmap
--model_weights	required	Path to weights directory
--output_dir	./outputs	Output directory
--num_inference_steps	1	Diffusion steps (1 is fast and works well)
--device	cuda	cuda or cpu
--torch_dtype	bfloat16	bfloat16 or float32

Python API

import numpy as np
from PIL import Image
from src.synclight.inference import evaluate, get_model

model = get_model("ckpt/", device="cuda")

images = [Image.open(p).convert("RGB") for p in image_paths]
lightmap = np.load("lightmap.npy").astype("float32")  # (H, W, 4)

outputs = evaluate(model, images, lightmap_image=lightmap, num_sampling_steps=1)
# outputs: list of PIL Images, one per input view

Repository Structure

synclight/
├── src/synclight/          # Core library
│   ├── models/
│   │   ├── lbm/            # LBM diffusion model
│   │   ├── unets/          # MVDream UNet with 3D attention
│   │   ├── vae/            # VAE wrapper
│   │   └── embedders/      # Conditioners (lightmap concat)
│   └── inference/          # Inference utilities
├── ckpt/
│   └── config.yaml         # Model config
├── input_images/           # Example inputs
├── outputs/                # Default output directory
├── gradio_demo.py          # Interactive web demo
├── lightmap_creator.py     # Interactive lightmap editor
└── test.py                 # Command-line inference script

Citation

@article{serrano2025synclight,
  title={SyncLight: Single-Edit Multi-View Relighting},
  author={Serrano-Lozano, David and Bhattad, Anand and Herranz, Luis and Lalonde, Jean-Fran\c{c}ois and Vazquez-Corral, Javier},
  journal={arXiv preprint arXiv:2601.16981},
  year={2025}
}

Acknowledgements

This work builds on the Latent Bridge Matching framework by Jasper AI and the MVDream multi-view diffusion architecture.

License

Apache 2.0