Head to head
A
B
Table (click a header to sort; click a row to scroll to its card)
Clips
Picks
Starred arms with their numbers, as markdown; copy this back into the session.
Scene: bakery (PUBLIC scene: 'exactly the same is a different question'; REAL checkpoints only - W4A8 ref vs NVFP4 all/nofc2/mid/fc1; 141 f T2VA, 25 steps, two lines, no cuts) (all scenes). Same seed everywhere; the reference is the exact W4A8 arm (ref_w4a8_exact). REAL CHECKPOINTS ONLY - nothing on this page is simulated. Each arm is a full weight file (built from the bf16 source) rendered on the native kernels: the shipped W4A8 reference, the shipped int8 convrot (1% weight error, the quality ceiling of the family), and four NVFP4 builds (all / nofc2 / mid / fc1; the mixes keep their unconverted layers bf16, which drags their wall time - a production mix would keep W4A8 there). Wall times are the regimes' actual speed; the VRAM column is a device-level nvidia-smi peak from a single run and includes the model-swap transient (it tracks checkpoint file size, not steady-state need; deltas under 2.5 GiB are not claims). Key result: the int8 control lands in the same distance band as full NVFP4 and moves the audio MORE - the band is take-noise, and no ruler orders quality inside this family.
Click a clip to play it with sound; hover plays muted; d on a card swaps in the amplified |arm - ref| view. Star candidates, send two to A/B, flip with f, choose which side you hear, wipe with the slider. Cost columns marked [PROJ] are projections from GEMM microbenches; everything else is measured.
Starred arms with their numbers, as markdown; copy this back into the session.