3 models, one chinchilla, $1.12: all three failed alike

The test started from a practical question: which animal scenes earn views. Counting 1,279 YouTube animal videos, the author found median views varied by an order of magnitude — 53.1 million for drinking scenes, 21.25 million for sneezing, but 5.16 million for grooming. Grooming looked safe and its motion easy to see, so it was chosen. That choice is what later mattered.

The first round changed only the model, holding image, prompt, length and judging consistent, with Codex scoring five items. All three models passed camera lock and centering and failed the other three, and their failure locations matched. wan-3.0, the most expensive at $0.10 per second, was the only one to fail centering and changed the face mid-clip. The two cheaper models matched item by item, so price did not predict quality.

The author then separated model limits from subject difficulty by hiding the front paws and shrinking the motion to whisker twitches and blinks, dropping wan-3.0. Fusion, warping and disappearance did not reappear. What remained across both remaining models were thin lines — whisker count and length, and forehead and cheek stripes. These point to a shared weakness in current models, not a gap between them.