chenkangjie1123/VGGT-Diff-NVS-480x832

VGGT-Diff NVS 480x832

This is the full-resolution independent-view checkpoint used for VGGT-Diff's quantitative comparison with other novel-view synthesis baselines. It is the paper's controlled clean-1K setting, with 980 available training scenes.

Configuration

• Resolution: 480 x 832
• Task: independent-view novel-view synthesis
• Input: six posed source views and target cameras
• VAE: source and target views are encoded independently
• Geometry: frozen VGGT-Omega features with camera conditioning
• Training data: DL3DV clean-1K (980 available scenes)
• Training schedule: 147 epochs at 192 x 336 (approximately 18K optimizer steps), followed by 60 epochs at 480 x 832
• Full-resolution continuation: 7,380 optimizer steps
• Intended use: the full-resolution NVS setting used for baseline comparison

For a controlled comparison, the paper follows the data scale and optimization budget of the FrameCrafter setting. The released checkpoint reflects that fixed comparison budget rather than a fully saturated endpoint. The full-resolution training curve did not show a clear convergence plateau, so continued optimization is expected to improve performance further.

The paper's 6-to-1 evaluation used 50 diffusion steps and target-only early Omega CFG with scale 2.0.

Usage

python scripts/infer.py \
  --training-mode nvs \
  --checkpoint /path/to/vggt-diff.safetensors \
  --omega-checkpoint /path/to/vggt_omega_1b_512.pt \
  --height 480 --width 832

VGGT-Omega weights and Wan2.1 base-model files are separate dependencies and are not redistributed in this repository.

Links

• Hugging Face Paper
• arXiv
• Code
• Project page

License

Released under the non-commercial VGGT-Diff Research License. Third-party models and datasets remain subject to their own terms.