PAVG trajectory checkpoints

PAVG predicts physical 3D point trajectories for solids, gases and solid–gas interactions. A separate pretrained video generator uses the projected trajectories to synthesize the visible video.

Paper · Project · Code and installation

Files

File Purpose
model.safetensors Joint solid, gas and interaction trajectory prediction
stage1/solid/model.safetensors Solid trajectory prediction and Stage 2 training initialization
stage1/gas/model.safetensors Gas trajectory prediction and Stage 2 training initialization
configs/stage2.yaml Joint model configuration for inference
configs/stage1.yaml Shared Stage 1 model architecture, with the solid training template
inference_protocol.json Trajectory and video generation defaults

The Stage 2 model contains both phase branches and their interaction modules. Standalone Stage 2 prediction requires model.safetensors and its matching configuration. The Stage 1 files are available for phase-specific use and Stage 2 training initialization.

DaS, CogVideoX and other video-generator weights are obtained separately from their authors. This repository contains PAVG trajectory weights.

Download and run

Install the public PAVG code and its pinned PhysCtrl dependency as described in the code repository. Download this model repository from the code directory:

hf download irony2025/PAVG --local-dir weights/pavg

Predict the trajectories for the provided interaction example:

python scripts/run.py predict -- \
  --config weights/pavg/configs/stage2.yaml \
  --checkpoint weights/pavg/model.safetensors \
  --phase-inputs examples/interaction/phase_inputs.npz \
  --condition-input examples/interaction/condition_input.npz \
  --output outputs/example/prediction.npz

The code README describes projection, tracking-video rendering and video generation. For custom inputs and motion controls, see the code repository's inference guide.

For training, generate phase-specific configurations with scripts/configure.py and point Stage 2 initialization to the downloaded Stage 1 weights. The supplied Stage 1 configuration is the solid template.

Acknowledgments

PAVG builds on PhysCtrl, Diffusion as Shader, CogVideoX and Diffusers. Third-party code and pretrained models retain their respective terms.

License

The PAVG trajectory checkpoints and PAVG-authored configuration files in this repository are licensed under the Apache License 2.0. Third-party code and video-generator weights retain their respective terms; see the third-party dependency guide.

Downloads last month

-

Downloads are not tracked for this model. How to track
Safetensors
Model size
36.4M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Paper for irony2025/PAVG