Hi maintainers —
Sharing a small independent held-out result on THUDM/CogVideoX-5b-I2V (Physics-IQ-style ball-and-block-fall prompts), for transparency:
- Protocol frozen before generation (VIDEO_CHECKS + static_hover)
- n=9 (seeds 2–4 × left/center/right): flag rate 1.000, Wilson 95% CI [0.701, 1.000], all via
static_hover
- Discovery n=6 (seeds 0–1) was post-hoc and is labeled separately
- Real Physics-IQ footage stays quiet under the same profile
- Explicit limit: one failure mode / scene family — not a general quality ranking
Writeup: https://haga.mushoodhanif.com/article/sim-physics-consistency-v1#held-out
Lab: https://haga.mushoodhanif.com/lab/physicsiq
Questions for anyone who has looked at CogVideoX physics failures:
- Do you see freeze/hover as a known failure mode on fall/release scenes?
- Are there public seeds / scenarios you’d recommend as harder controls?
- Permission to score additional public demo clips under the same frozen detectors (aggregate-only publication)?
Happy to correct anything misstated about the model card / recommended settings. Thanks for releasing the weights.
Hi maintainers —
Sharing a small independent held-out result on THUDM/CogVideoX-5b-I2V (Physics-IQ-style ball-and-block-fall prompts), for transparency:
static_hoverWriteup: https://haga.mushoodhanif.com/article/sim-physics-consistency-v1#held-out
Lab: https://haga.mushoodhanif.com/lab/physicsiq
Questions for anyone who has looked at CogVideoX physics failures:
Happy to correct anything misstated about the model card / recommended settings. Thanks for releasing the weights.