# Example videos — VideoPhy-2 OpenVid ablation

Five prompts × five variants = 25 MP4s. Filename: `{id}_{variant}.mp4`. All videos generated at 49f / 720×480 / 8 fps / 50 steps / guidance 6.0 / seed 42 with the official VideoPhy-2 mPLUG-Owl autograder scoring SA (semantic adherence) and PC (physical commonsense), each on a 1-5 scale; Joint = SA≥4 AND PC≥4.

| id   | action                  | caption                                                                                        | base PC/SA | std PC/SA | xfm PC/SA | k1 PC/SA | k5 PC/SA |
|------|-------------------------|------------------------------------------------------------------------------------------------|-----------:|----------:|----------:|---------:|---------:|
| 0022 | ripping paper           | A hand rips a sheet of printer paper in half, creating a jagged tear.                          | 4 / 4 ✓    | 3 / 3     | 4 / 3     | 4 / 4 ✓  | 4 / 4 ✓  |
| 0305 | attaching to anchor     | A rock climber attaches a rope to a climbing anchor.                                           | 4 / 4 ✓    | 3 / 4     | 4 / 3     | 4 / 4 ✓  | 4 / 4 ✓  |
| 0472 | knitting (purl stitch)  | Close-up showing the process of creating a purl stitch, showing the movement of yarn/needle.   | 5 / 4 ✓    | 5 / 3     | 5 / 3     | 5 / 4 ✓  | 5 / 4 ✓  |
| 0006 | smoking (smoke rings)   | A person blows smoke rings into the air, the rings briefly visible before dissipating.         | 4 / 4 ✓    | 5 / 4 ✓   | 4 / 4 ✓   | 4 / 4 ✓  | 4 / 4 ✓  |
| 0245 | somersaulting (hard)    | A person does a somersault, rolling onto their back and immediately getting up.                | 2 / 3      | 2 / 3     | 2 / 3     | 2 / 3    | 2 / 3    |

✓ = Joint passing (SA≥4 AND PC≥4).

Selection rationale:
- **0022, 0305, 0472**: GDN-CSSM (k1+k5) recovers Joint where Standard-LoRA / Transformer-LoRA dropped SA below the threshold. These are the "ablation matters" cases.
- **0006**: every variant passes — sanity check that the LoRAs don't degrade clean prompts.
- **0245**: every variant fails (PC=2). Classic CogVideoX-5B failure mode — articulated body inversion, no LoRA at rank=16 fixes it. Motivates the rank=128 follow-up.