README.md

metadata

title: WildRayZer
emoji: 🎥
colorFrom: indigo
colorTo: gray
sdk: gradio
app_file: app.py
pinned: false
license: other

WildRayZer — Gradio Space

Demo for WildRayZer: Self-supervised Large View Synthesis in Dynamic Environments (CVPR 2026, Highlight).

On first launch the Space downloads the 2-view checkpoint from uva-cv-lab/wildrayzer-2view via huggingface_hub.hf_hub_download. DINOv3 ViT-7B is also fetched from the hub on first model construction (~30 GB). Both downloads are cached under HF_HOME — set it to the persistent storage mount in your Space settings.

Hardware

Requires a CUDA GPU with ≥ 40 GB VRAM (DINOv3 ViT-7B + scene encoder + renderer). CPU Spaces will not work.

Files

• app.py — Gradio UI (scene picker, threshold slider, video toggle).
• app_demo.py — WildRayZerDemo class: model loading, batching, rendering.
• configs/wildrayzer_inference.yaml — inference config.
• model/ — rayzer_official_v3.py + transformer.py + loss.py.
• utils/ — supporting utilities (camera, metrics, data, positional embeddings).
• data/dynamic_re10k/test/ — a subset of the D-RE10K test split (6 scenes + the filtered view-index JSON). Full dataset at https://huggingface.co/datasets/uva-cv-lab/Dynamic-RE10K.
• setup.py — config loader / DDP helpers imported by inference.py (kept for cross-compatibility).

Env vars

• WILDRAYZER_CKPT — local path to a checkpoint, overriding the HF download.
• WILDRAYZER_CFG — local path to a config YAML.
• WILDRAYZER_TEST_ROOT — path to a D-RE10K-style test directory.
• HF_HOME — cache dir for checkpoint + DINOv3; point at /data on paid Spaces.