ovie-ft-re10k

Metrics

Values reported in the paper for this checkpoint. RealEstate10K test split, 750 scenes, stride 3, 14 target frames (in-domain for this checkpoint):

Fine-tuning selects a domain: on DL3DV, where this checkpoint is out-of-domain, the base model is stronger on pixel fidelity, perceptual similarity and FID (PSNR 14.8 vs 14.1, FID 13.6 vs 24.66), while this checkpoint keeps the better multi-view consistency (MEt3R 0.040 vs 0.078). For comparison, the same architecture trained from scratch on RealEstate10K alone reaches only 19.05 PSNR, 2.85 dB below this fine-tune — the gap comes from monocular pretraining.

Usage

import torch
from models.models import OVIEModel
from utils.pose_enc import extri_intri_to_pose_encoding
from torchvision.transforms import ToTensor
from PIL import Image

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

model = OVIEModel.from_pretrained("kyutai/ovie-ft-re10k").to(device)
model.eval()
image_size = model.image_size  # 256

img_pil = Image.open("image.jpg").convert("RGB").resize((image_size, image_size))
img_tensor = ToTensor()(img_pil).unsqueeze(0).to(device)

extrinsics = torch.tensor([[[1.0, 0.0, 0.0, -1.25],
                            [0.0, 1.0, 0.0,  0.5],
                            [0.0, 0.0, 1.0, -2.0]]], device=device)
dummy_intrinsics = torch.zeros(1, 1, 3, 3, device=device)

camera = extri_intri_to_pose_encoding(
    extrinsics=extrinsics.unsqueeze(0),
    intrinsics=dummy_intrinsics,
    image_size_hw=(image_size, image_size),
)
cam_token = camera[..., :7].squeeze(0)

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

To reproduce the metrics with the repository's evaluation script:

uv run python evaluate.py \
    --dataset_path /PATH/TO/RE10K/TEST \
    --config_path configs/config_ovie.yaml \
    --from_pretrained kyutai/ovie-ft-re10k \
    --stride 3 --num_target_frames 14

See the repository for installation, data preprocessing, 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},
}