ovie-ft-dl3dv

In-the-Wild Monocular Pretraining for Novel View Generation

Part of the OVIE collection.

This is kyutai/ovie fine-tuned for 50k steps on DL3DV, the companion to the RealEstate10K fine-tune. The architecture is identical to the base model; only the weights differ.

It is the control that shows fine-tuning selects a domain: applying the same protocol on DL3DV markedly improves in-domain reconstruction relative to the base model, but costs distributional realism and generalization.

Metrics

Fine-tuning raises DL3DV PSNR by 2.3 dB while raising FID from 13.6 to 17.24: the model becomes a more faithful per-view reconstructor at a slight cost to global distributional realism. Notably it retains — and on RealEstate10K slightly improves — cross-domain performance even though that benchmark is now out-of-domain (PSNR 19.34 vs 18.8), so the fine-tune sharpens in-domain fidelity without eroding generalization here.

Prefer the base OVIE for in-the-wild zero-shot use and the RealEstate10K fine-tune for indoor real-estate scenes. This checkpoint is intended for DL3DV-style captures.

Usage

import torch
from models.models import OVIEModel

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = OVIEModel.from_pretrained("kyutai/ovie-ft-dl3dv").to(device)
model.eval()
image_size = model.image_size  # 256

The forward pass is identical to the base model:

with torch.no_grad():
    pred = model(x=img_tensor, cam_params=cam_token)  # (1, 3, 256, 256) in [0, 1]

See the base model card for the complete pose-encoding snippet, and the repository for installation and the full benchmark table.

Citation

@misc{ovie2026,
      title={One View Is Enough: In-the-Wild Monocular Pretraining for Novel View Generation},
      author={Adrien Ramanana Rahary and Nicolas Dufour and Patrick Perez and David Picard},
      year={2026},
      eprint={2603.23488},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.23488},
}