The ask
prompts, six clips each, ~8 minutes. For every prompt: pick the one you would ship, optionally pick the best-sounding one, and flag anything you would not ship at all. Labels stay hidden until you choose to reveal them.
Judge each clip on whether you would ship it, never on whether it matches the others. Same seed, different schedules — they walk different denoise paths, so they show different people and rooms. That is a free choice, not a defect. "Closest to the baseline" has already fooled two metrics on this model.
Clips have sound. This model generates video and stereo audio in one pass, and the audio is the least-tested part. The grid plays muted — click any clip to open it with sound.