28 prompts × 2 arms, 1344×768, 241 frames (10 s @ 24 fps), seed 4242, one RTX 5090. Every clip generated by POSTing to a real image-server instance — the production path, not a bench script. The arms differ in weights only: same code, same sampler, same seed. We ship INT8; BF16 is here purely as the unquantized reference.
Clips carry audio — 7 prompts contain dialogue, and audio is part of what quantization can damage, so play them with sound. Only one clip plays at a time.
PSNR/SSIM are INT8 measured against BF16, and they are a divergence metric, not a quality score. At 8 steps with CFG=1 a small weight perturbation can tip the sampler into a different mode: 21 of 28 prompts sit in 22–30 dB (visually near-identical), but 3 fall below 20 dB because the arms rendered different content. The worst, 05_product_perfume at 14.6 dB, is a camera orbiting the opposite way — and it is INT8 that follows the prompt's “arcs a short way to the right”. Do not read a low score as the quantized arm being worse.
Metrics were computed on the encoded H.264, so a little of the number is codec; both arms were encoded identically. BF16 ran at a lower mmgp budget (0.4% vs 0.65%) purely to stay clear of the 31.4 GiB ceiling — that affects its speed, not its frames, so do not read a speed ratio off this page. Matched-budget timings at 1280×704×241 are INT8 43 s vs BF16 68 s.