[ECCV 2026] MetaView: Monocular Novel View Synthesis with Scale-Aware Implicit Geometry Priors

Yufei Cai¹  Xuesong Niu^2*  Hao Lu³  Kun Gai²  Kai Wu^2†  Guosheng Lin^1†

¹ Nanyang Technological University  ·  ² Kuaishou Technology  · 

³ The Hong Kong University of Science and Technology (Guangzhou)

^* Project lead    ^† Corresponding author

   

Demo

🚀🚀 Interactive Demo: Try our demo in the Huggingface Spaces!

⚡️⚡️ Fast Inference: Try 8-step distilled lightning demo built with an 8-step distilled lightning model for fast inference!

🤗🤗 Acknowledgement: Special thanks to Apolinario and the Huggingface open-source team for their support and help with this project!

Overview

MetaView is a diffusion-based framework for high-fidelity novel view synthesis that enables accurate rendering under large view changes from a single image.

Current generative novel view synthesis methods typically rely on restrictive explicit 3D reconstruction pipelines to enforce spatial consistency but inherently restrict generalization in large view changes. Morever, recent interactive generative methods suffers from scale drifting and poor geometry consistency. MetaView bridges this gap by combining implicit geometry modeling with minimal yet essential explicit 3D cues:

• Scale-Aware Implicit Geometry Priors: We incorporate implicit geometry priors from a feed-forward geometry perception network (Depth Anything 3) to regularize structure without imposing restrictive reconstruction pipelines. These geometric signals are incorporated into the pretrained MM-DiT backbone via non-invasive parallel attention layers.

• Metric Scale Anchoring via Modified RoPE: To overcome the scale drifting issue prevalent in fully implicit methods, we allocate an extra subspace for the z-axis to anchor the scene scales and it plays with camera pose via PRoPE. This explicitly injects scale cues, anchoring the generation to a consistent 3D metric space.

Given a single input image and a target camera pose, MetaView synthesizes the corresponding novel view with precise camera controllability, strong geometry consistency, and remarkable cross-domain generalization.

Environment Setup

1. Inference Environment: metaview

Create the metaview conda environment with dependencies for DiffSynth-Studio and Depth Anything 3:

git clone https://github.com/KlingAIResearch/MetaView.git
cd MetaView

conda create -n metaview python=3.10 -y
conda activate metaview

pip install -e .

# Additional dependencies for inference
pip install opencv-python pillow openexr_numpy

# Install Depth Anything 3
git clone https://github.com/ByteDance-Seed/Depth-Anything-3.git
cd DepthAnything3

pip install -e .

Note on DiffSynth-Studio Installation: The diffsynth/ source code and pyproject.toml are included in this package, so pip install -e . will install DiffSynth-Studio and its dependencies (transformers, accelerate, safetensors, peft, sentencepiece, etc.) following the official DiffSynth-Studio installation guide.

Note on Depth Anything 3: Installed from the local DepthAnything3/ source following the official DA3 installation guide.

2. Evaluation Environment: metric

Create the metric conda environment with UFM dependencies for running evaluation metrics:

cd MetaView
git clone --recursive https://github.com/UniFlowMatch/UFM.git
cd UFM

conda create -n metric python=3.10 -y
conda activate metric

# Install UniCeption (UFM submodule dependency)
cd UFM/UniCeption
pip install -e .

# Install UFM
cd ..
pip install -e .

# Additional dependencies for metric evaluation
pip install lpips flow_vis scikit-image scikit-learn matplotlib tqdm

Model Weights

Download the following model weights and place them in models/:

Model	Usage	Download
MetaView	Pretrained checkpoint	MetaView
DA3-GIANT-1.1	3D feature extraction	Depth-Anything-3-GIANT
DA3NESTED-GIANT-LARGE-1.1	Dense depth estimation	Depth-Anything-3-NESTED
Qwen-Image-Edit	Base DiT model (auto-downloaded)	Qwen/Qwen-Image-Edit

Recommended Directory Layout

models/
├── Qwen/
│   └── Qwen-Image-Edit/
├── MetaView/
│   └── checkpoint-best.safetensors    # MetaView pretrained checkpoint
└── Depth-Anything-3/
    ├── DA3-GIANT-1.1/
    └── DA3NESTED-GIANT-LARGE-1.1/

Usage

Inference

Modify the infer.sh (or src/inference.py directly) to generate a novel view from a single image:

conda activate metaview

# Edit paths inside before running
bash infer.sh

Or run with full CLI arguments:

python src/inference.py \
    --image_path "./assets/data/0.png" \ # Try samples in `assets/data/` to reproduce or specify your input image.
    --ckpt_path "Path/to/MetaView/checkpoint" \
    --da3_giant_path "Path/to/DA3-GIANT/model" \
    --da3_depth_path "Path/to/DA3-NESTED/model" \
    --yaw -30 \
    --pitch 15 \
    --radius 3.5 \ # if None, auto-calculated from center depth
    --output_path "./result/output_novel_view.png"

Arguments

Argument	Default	Description
--image_path	(required)	Path to the input image
--ckpt_path	(required)	Path to the pretrained MetaView checkpoint (.safetensors)
--output_path	./output_novel_view.png	Path to save the generated image
--yaw	0.0	Yaw angle in degrees (positive = right, negative = left)
--pitch	0.0	Pitch angle in degrees (positive = up, negative = down)
--radius	None (auto)	Rotation radius. If None, auto-calculated from center depth
--da3_giant_path	(required)	Path to DA3 GIANT model
--da3_depth_path	(required)	Path to DA3 depth model
--qwen_path	None (auto)	Base path to Qwen-Image-Edit model

The output is a side-by-side image: [Source Image | Generated Novel View].

Evaluation Metrics

Modify eval.sh in the metric environment to compute PSNR, SSIM, LPIPS, and proposed DMD:

conda activate metric

# Edit paths inside before running
bash eval.sh

Or run with full CLI arguments:

python -u src/metric.py \
    --data_path "Path/to/evaluation/folder" \
    > ./eval_output.txt  2>&1 

The metric script expects input images to be 3-panel stitched images: [Source Image | Generated | Ground Truth], each panel being 960×528.

Metrics reported:

• PSNR (Peak Signal Noise Ratio)
• SSIM (Structural Similarity)
• LPIPS (Learned Perceptual Image Patch Similarity, VGG backbone)
• DMD (proposed Dense Matching Distance)

Citation

@article{metaview,
      title={MetaView: Monocular Novel View Synthesis with Scale-Aware Implicit Geometry Priors}, 
      author={Yufei Cai and Xuesong Niu and Hao Lu and Kun Gai and Kai Wu and Guosheng Lin},
      year={2026},
      eprint={2607.12000},
      archivePrefix={arXiv},
      url={https://arxiv.org/abs/2607.12000}, 
}

Acknowledgements

This codebase builds upon:

• DiffSynth-Studio
• DepthAnything3
• UFM