ViewRecNet

ViewRec3D: Learning to Recommend 3D Viewpoints for AI Photography

Chengzhi Zhang^1,*, Yunzhong Hou^1,*,†, Zizheng Sun¹, Ming-Hsuan Yang², Chi Harold Liu¹

¹ Beijing Institute of Technology   ² University of California, Merced

^* Equal contribution.   ^† Corresponding author.

Overview

Given a single photograph, ViewRecNet recommends how to move and rotate the camera, adjust the field of view, and choose portrait or landscape orientation to approach an expert composition. The model uses a VGGT backbone and learns from paired views in ViewRecDB-100K with dense camera-ray supervision.

Released checkpoint

model.safetensors (4.51 GB) contains the weights from training step 15,625, including the backbone, camera head, orientation head, and ray head. The evaluation code constructs the network and loads these weights directly.

Quick start

From the ViewRec3D code repository, install the project requirements and download the checkpoint:

python -m pip install -r requirements.txt
python -m pip install -U huggingface_hub
hf download czZhang2026/ViewRecNet model.safetensors --local-dir checkpoints/ViewRecNet

To evaluate on a CUDA GPU, prepare the test split as described in the dataset card. Replace the test directory below with its local path. The evaluator reads the checkpoint from --output_dir and writes evaluation metrics there.

python -m src.evaluate --val_dir /path/to/ViewRecDB-100K/test --output_dir checkpoints/ViewRecNet --use_rays True

Acknowledgements

This work was supported by the National Natural Science Foundation of China (No. U25A20434).