Modelpublic

JohnZhan/MotionInsight-8B

sourceHugging Faceapache-2.0updated 23d agoView on Hugging Face

1likes40downloads

Model Card

MotionInsight-8B

This is the released multidim-all/checkpoint-5500 for MotionInsight: Diagnosing Object Motion Deficiencies in Generated Videos.

The model consumes sampled RGB frames, target-object motion features from SAM3 and CoTracker3, and camera motion from VIPE. It predicts diagnostic reasoning and three motion-quality scores.

Use this checkpoint with the code and custom Qwen3-VL implementation from:

https://github.com/JohnZhan2023/MotionInsight

bash

hf download JohnZhan/MotionInsight-8B \
  --local-dir checkpoints/MotionInsight-8B
python scripts/patch_transformers.py
python inference.py --help

The VidMotion dataset, videos, and annotations are not distributed.