GeoNVS: Geometry Grounded Video Diffusion for Novel View Synthesis

GS-Adapter and LoRA weights for GeoNVS, which injects explicit 3D Gaussian geometry into camera-controlled video diffusion models.

Project page  |  Paper  |  Code

These weights are fine-tuned from stabilityai/stable-virtual-camera (SEVA) and CameraCtrl (SVD backbone), and inherit the Stability AI Non-Commercial Research Community License — using them requires the base models and acceptance of their licenses. Non-commercial research use only.

Files

Path	Backbone	Description
eccv_gattn/	SEVA	Main GeoNVS model, Adaptive Fusion (fusion_method: gattn): gs_adapter_weights.pth + pytorch_lora_weights.safetensors
eccv_camctrl/	CameraCtrl (SVD)	GeoNVS on CameraCtrl: adapter + LoRA
eccv_base/	SEVA	Fusion ablation: Naive Fusion (fusion_method: concat), use with configs/module_config/gsadapter_eccv_base.yaml

Usage

Use with the GeoNVS code release — tools/scripts/download_weights_ours.sh fetches these files into pretrained_weights/, then:

python demo.py --data_path <benchmarkset>/dl3dv10 --num_inputs 3 \
    --gs_adapter_config configs/module_config/gsadapter_eccv_gattn.yaml \
    --gs_adapter_weight_path pretrained_weights/eccv_gattn \
    --lrm_model_name vggt_iv --H=384 --W=384 --version=1.0

Mirrored third-party checkpoint

third_party/depthsplat-gs-base-re10kdl3dv-448x768-randview2-6-f40abc4f.pth is redistributed from DepthSplat (MIT License, © 2024 Haofei Xu). It is the checkpoint used for the paper's experiments; upstream has since replaced it with a retrained file (f8ddd845) that measurably changes results, so the original is mirrored here for reproducibility. Its MIT license governs that file.