Pass 1: what the solver does to the movie
Thirteen takes of one scene. Same prompt, same seed, same 1024×1024, same 124 frames, same checkpoint. The only change is the pass-1 sampler, scheduler and step count, and there is no second pass anywhere. Which ones read best to you?
Why we ran it: we froze a fresh pass 1 for a pass-2 sampler study and the clips came back visibly unlike the scene we bench against. Not a variation on it, a different take. So we measured it.
The control that makes the rest readable
Re-running the exact recipe behind our shipped clip today, three weeks and a core update later, reproduces it at 39.9 dB, tighter than that clip's own re-encode at 37.8 dB. The pipeline is reproducible, so every difference below is the setting and not drift in the stack. Every other take lands 14-18 dB from the reference, further apart than two different pass-2 samplers on an identical pass 1 (23-26 dB). There seems to be no small change to pass 1: touch any part of the solve and you get a different movie rather than a nudged one.
Read the clips, not the table
drift is distance from the reference take, not quality.
motion is mean frame-to-frame luminance difference, and on a lateral pan
that is dominated by the camera, so it ranks gross energy and cannot tell you whether
a character's turn reads well. That is the part we want eyes on. We have our own
ordering and are deliberately not stating it so as not to anchor yours.
Click a row to load it against the reference. One pair plays at a time, posters until you ask.
What it does not explain
Drift does not track preference: in our own viewing the two furthest-drifting takes
sit at opposite ends of the ranking. Step count is not uniform either, and even changes
sign: res_multistep + simple and er_sde + beta
gain gross motion as steps rise, while both gradient_estimation recipes lose
it. One scene, one seed, so treat all of it as a starting point.