VGGT-Diff NVS 192x336 (100K)
This is the half-resolution independent-view scaling checkpoint trained on the DL3DV clean-10K split (9,510 available scenes) for 100,000 effective optimizer steps. It takes six posed source images and jointly predicts one or more target views.
Configuration
- Resolution: 192 x 336
- Task: independent-view novel-view synthesis
- Input: six posed source views and one or more target cameras
- VAE: source and target views are encoded independently
- Geometry: frozen VGGT-Omega features with camera conditioning
- Training data: DL3DV clean-10K (9,510 available scenes)
- Training: full-model fine-tuning
- Release checkpoint: 100K effective training steps
This release extends the original 18K fixed-compute setting to 100K steps. Although individual checkpoints can fluctuate, the long-term evaluation trend remains favorable and the 100K checkpoint is best across the reported half-resolution image metrics. PSNR is approaching a plateau, while SSIM and LPIPS continue to improve, so this release should not be interpreted as a fully saturated endpoint.
Usage
Use the nvs architecture mode:
python scripts/infer.py \
--training-mode nvs \
--checkpoint /path/to/vggt-diff.safetensors \
--omega-checkpoint /path/to/vggt_omega_1b_512.pt \
--height 192 --width 336
VGGT-Omega weights and Wan2.1 base-model files are separate dependencies and are not redistributed in this repository.
Links
License
Released under the non-commercial VGGT-Diff Research License. Third-party models and datasets remain subject to their own terms.