DeblurNVS: Geometric Latent Diffusion for Novel View Synthesis from Motion-Blurred Images

Weights | Project Page | ArXiv

A local demo release for our three-stage novel view synthesis pipeline.
Given only motion-blurred multi-view input images, the system estimates DA3-native cameras, restores sharp context latents, synthesizes target-view latents along an interpolated trajectory, and decodes them into RGB novel views.

Checklist

• Release the demo code and pretrained weights.
• Release the full dataset and evaluation scripts.
• Release the training code.

Quick Start

1. Install dependencies

Use your existing environment and install the required Python packages:

pip install -r requirements.txt

2. Prepare weights

Download or sync the released runtime assets into pretrained/.

Weights are hosted at:

https://huggingface.co/ChangyueShi/DeblurNVS

After downloading, the pretrained/ directory should look like:

pretrained/
├── da3_base/
│   ├── config.json
│   └── model.safetensors
├── normalization_stats_level1.pt
├── stage1_lora.pt
├── stage2_diffusion.pt
└── stage3_decoder.pt

Repository Layout

deblurnvs_opensource/
├── deblurnvs/          # demo runtime
├── example/            # bundled input-only example scene
├── pretrained/         # all local runtime assets
├── utils/              # local helper code and lightweight model components
├── requirements.txt
└── run_demo.py

Input Format

Each scene should contain blurred input images in one of the following layouts:

scene_root/
└── images_train/
    ├── 000.png
    ├── 001.png
    └── ...

Run the Example

python run_demo.py \
  --scene-root example \
  --context-views 9 \
  --num-novel-views 25 \
  --output-dir outputs/demo_interp \
  --device cuda:0

Output Format

outputs/demo_interp/
├── context_views/
├── context_views_pred/
├── pred/
├── overview.png
├── camera_path.json
└── metadata.json

Output	Description
context_views/	input blurred context images
context_views_pred/	reconstructed RGB predictions for observed views
pred/	interpolated novel-view RGB predictions
overview.png	compact visual summary
camera_path.json	exported interpolated target-camera path
metadata.json	run configuration and checkpoint metadata

Citation

This is a demo repository. Some implementation details may differ slightly from the final paper version.