PHYSICS-AWARE VIDEO GENERATION & REPAIR

PhysLoop:

Harnessing LLMs' Physics Awareness for Self-Improving Video Generation Models

1Peking University2ByteDance Seed3University of California, Santa Barbara4Massachusetts Institute of Technology5Universität Hamburg6Hangzhou Normal University

* Equal contributions · ‡ Industry advisor · † Project lead · § Corresponding author

VideoCodeDatasetsPaperForthcoming

Overview

PhysLoop workflow for training-data construction and video repair.

Dataset Examples

Benchmark Results

MethodHuman (Gen)Model (Gen)Task
Phys.Real.CausalWinPCSAIQAQGenRepairP-IQ
Video Generation Models
Gemini Omni3.113.823.1818.9420.7939.1771.1750.4049.0014.6731.66
H33.634.273.5544.9322.9838.6868.4750.6253.0018.0028.14
Seedance 2.53.804.293.8053.7921.5138.6967.6350.6373.0019.3348.07
Seedance 2.5
Claude Opus 53.823.823.8464.1820.2336.6967.6649.7171.0078.6753.75
Kimi K33.633.693.6954.4119.9037.3268.5450.4269.0068.6752.04
GPT-6 Astra
Gemini Omni3.532.963.5145.4512.7135.8475.1146.7267.0062.0012.08
H33.863.453.9058.2117.4837.1268.7748.6573.0072.6739.10
w/o Refiner3.942.593.9534.3312.5334.8872.4245.7580.0060.67—
GPT-6 Astra + Seedance 2.5
PhysLoop4.063.724.1275.7615.2936.2869.9348.7085.0086.0052.06

BibTeX

Dataset

ComponentProgressStatusHostingNotes
PhysLoop-GenPreparingNot releasedHugging Face (planned)Generation evaluation; inputs and protocol under review
PhysLoop-RepairInventory pendingNot releasedHugging Face (planned)Repair pairs and evaluation protocol under review
PhysLoop-100KCounts pendingNot releasedHugging Face (planned)Planned training corpus; accepted counts to be verified
Code & release updates on GitHub ↗

Community

WeChat

WeChat contact QR code

Add with PhysLoop for a group invite.