Answers are grounded in this paper and cite their sources.
Kinematic acceptance tests for post-training quantization of human-video diffusion transformers assume that low-bit arithmetic degrades generated motion.This work tests the assumption on DynamiCtrl at W8A8 and W4A8, with refinement of the driving pose and of the extracted skeleton.Every comparison is referred to the difference between two bfloat16 renders that differ only in the seed, with inference at the clip level.No difference from the bfloat16 reference was detected in any extracted skeleton statistic, whereas pose error and pixel distance were smaller than between seeds; equivalence within half of the seed-to-seed change was not established.Calibrating the quantizer on other subjects left pose and pixel error unchanged.With fixed thresholds the kinematic gate rejected 11 of 12 seed pairs; with calibrated thresholds, 21 of 24 W8A8 and 20 of 24 W4A8 videos passed, against 9 of 12 seed pairs, rates that depend on the number of calibration clips.
No takes yet. Share an insight, caveat, or question.
Albaghdadi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: