Calibrating a Biomechanical Acceptance Test for Post-Training Quantization of a Video-Diffusion Transformer

Abstract

Post-training quantization (PTQ) compresses large video-diffusion transformers without fine-tuning, and acceptance tests built from biomechanical measurements of the generated motion have been proposed as an alternative to frame-averaged reconstruction quality. This study examines whether such a test separates quantized from unquantized renders once it is calibrated against the numerical variability of the generator itself. For DynamiCtrl, a 5.77-billion-parameter pose-conditioned video-diffusion transformer, a null distribution is constructed from 57 pairs of numerically equivalent bfloat16 renders that differ only in attention kernel, weight dither at the scale of one unit in the last place, or execution platform. Seven stipulated thresholds accept only 14 of the 57 null pairs. Recalibrated to a 5% null rejection rate, the test accepts 75–97% of the renders in every W8A8 configuration examined, including full-model quantization on 38 of 39 renders; after multiplicity correction no check separates any configuration from the null pair by pair (smallest adjusted p = 0.088), and a clip-level paired analysis separates one of 17 configurations. PSNR against the same-seed bfloat16 render separates the full-model and blocks-1–13 configurations from the null: full W8A8 lies 2.4–3.0 dB below the null median of 34.2 dB, quantizing only blocks 1–13 reproduces most of that deficit, and two hybrid maps that keep those blocks in bfloat16 return to within 0.9 dB of the null while reducing DiT weight memory from 11.54 to 8.498.89 GB. Acceptance tests of this kind therefore require null calibration before their verdicts can be interpreted.