nyu-visionx/pisa-experiments

Pisa Experiments

This repository contains the PisaBench, training data, model checkpoints, introduced in PISA Experiments: Exploring Physics Post-Training for Video Diffusion Models by Watching Stuff Drop.

PisaBench

Real World Videos

We curate a dataset comprising 361 videos demonstrating the dropping task.Each video begins with an object suspended by an invisible wire in the first frame. We cut the video clips to begin as soon as the wire is released and record the videos in slow-motion at 120 frames per second (fps) with cellphone cameras mounted on tripods to eliminate camera motion.

We save each video in the following fomat:

├── 00000.jpg
├── 00001.jpg
...
├── movie.mp4
└── clip_info.json

• clip_info.json is a json file that contains positive/negative point annotations and text descriptions for each video.

Real world videos can be found at: pisabench/real.zip.

Simulated Test Videos

Since our post-training process uses a dataset of simulated videos, we also create a simulation test-set of 60 videos for understanding sim2real transfer. We create two splits of 30 videos each: one featuring objects and backgrounds seen during training, and the other featuring unseen objects and backgrounds.

We save each video in the following format:

├── rbga_00000.jpg
├── rbga_00001.jpg
...
├── movie.gif
├── mask.npz
├── clip_info.json

• mask.npz is segmentation masks for all objects with shape [V, N, H, W], where V is the number of video frames, N is the number of objects, H is the height, and W is the width.
• clip_info.json is a json file that contains annotations and text descriptions for each video.

Simulated test videos can be found at: pisabench/sim.zip.

Training Data

We use Google's Kubric for generating simulated physics videos. Kubric combines PyBullet and Blender for handling simulation and rendering seamlessly in a unified library.

We use the Google Scanned Objects (GSO) dataset which is already supported in Kubric. The GSO dataset consists of ~1000 high quality 3D objects that come from scans of a variety of everyday objects.

Training data can be found at:

• Physics Supervised Fine-Tuning (PSFT): training_data/psft.zip.
• Object Reward Optimization (ORO): training_data/oro.zip.

Checkpoints

Our approach for post-training is inspired by the two-stage pipeline consisting of supervised fine-tuning followed by reward modeling commonly used in LLMs.

Checkpoints can be found at:

• Open-Sora + PSFT (base): /checkpoints/base.
• base + ORO (Seg): /checkpoints/oro_seg.
• base + ORO (Flow): /checkpoints/oro_flow.
• base + ORO (Depth): /checkpoints/oro_depth.

Download Dataset

PisaBench can be downloaded using the following code:

from huggingface_hub import snapshot_download

dataset_path = 'PATH'   # The local directory to save downloaded dataset
snapshot_download("nyu-visionx/pisa-experiments", local_dir=dataset_path, repo_type='dataset')

Citation