NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis

Pretrained weights for NAMVIS (NeurIPS 2026), a diffusion-free, geometry-conditioned next-scale autoregressive model for sparse-view novel view synthesis.

• 📄 Paper: arXiv:2610.04722
• 🌐 Project page: https://corl-team.github.io/namvis
• 💻 Code: https://github.com/corl-team/namvis
• 🗂️ Training data: Part 1 · Part 2

Files

File	Description
namvis_1b.pth	NAMVIS 1B transformer (256×256)
infinity_vae_d32reg.pth	Multi-scale visual tokenizer from Infinity (frozen, unchanged, MIT)

Usage

hf download smileyenot983/NAMVIS namvis_1b.pth infinity_vae_d32reg.pth --local-dir weights

Inference, evaluation, and fine-tuning commands are in the code repository.

Details

• Resolution: 256×256
• Input: one or more posed source images and target camera poses
• Trained on renders of Objaverse-XL objects: Sketchfab objects (Part 1) and GitHub objects with an object-quality score above 6.5 (Part 2). See Objaverse-XL for per-asset licenses.

Citation

@inproceedings{khafizov2026namvis,
  title         = {{NAMVIS}: Next-Scale Autoregressive Multi-View Image Synthesis},
  author        = {Khafizov, Ramil and Statsenko, Ilya and Rakhimov, Ruslan and Komarichev, Artem and Wonka, Peter and Burnaev, Evgeny},
  booktitle     = {Advances in Neural Information Processing Systems},
  year          = {2026},
  eprint        = {2610.04722},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.04722}
}