Skip to content

Cosmos3 PhysicsIQ Reproduction

End-to-end recipe for reproducing the PhysicsIQ benchmark with Cosmos3-Super using the native Cosmos Framework PyTorch entrypoint (python -m cosmos_framework.scripts.inference).

The notebook walks through both PhysicsIQ task formats:

  • Image-to-Video (I2V): condition on the single switch-frame JPG; generate 121 frames (1 conditioning + 120 generated); score the last 120.
  • Video-to-Video (V2V): condition on a 3-second clip; pad the official 72-frame conditioning to 73 frames; generate 193 frames total; score frames 73..192.

Files

  • run_with_cosmos_framework.ipynb — main notebook.
  • assets/i2v_prompts.json — 198 I2V upsampled prompts.
  • assets/v2v_prompts.json — 198 V2V upsampled prompts.

Reference scores (Cosmos3-Super)

Task PhysicsIQ score
I2V 43.8
V2V 59.7

Requirements

  • 4-GPU Linux node (configurable via COSMOS3_NUM_GPUS, default 4)
  • uv == 0.11.29
  • ffmpeg, git, gcloud (or gsutil)
  • Hugging Face access to the Cosmos3 model family