VolFill v1.0 — Inference Release

First public release: inference code and pretrained checkpoints for VolFill (single-view amodal 3D scene reconstruction with volumetric flow matching).

Included

• End-to-end inference pipeline (image → MoGe visible geometry → visible TUDF → VAE encode → flow-matching DiT → VAE decode → 256³ TUDF).
• Pretrained VAE + DiT checkpoints — auto-download via --hf_repo TuanNgo/VolFill / from_pretrained, or manual from Hugging Face / Google Drive.
• Point-cloud visualizer (volfill.visualize, viser) and 3 demo samples.

Not included (coming soon)

Training pipeline, evaluation scripts (SCRREAM, NRGB-D), and dataset preprocessing.

Project page: https://ngoductuanlhp.github.io/VolFill/ · Paper: https://arxiv.org/abs/2605.31466