Q-Drift: Quantization-Aware Drift Correction for Diffusion Model Sampling

Name: Q-Drift: Quantization-Aware Drift Correction for Diffusion Model Sampling
Uploaded: 2026-10-04T01:34:29.856824+00:00
Duration: 363 s
Description: Q-Drift corrects generation quality lost due to quantization noise accumulating during diffusion model sampling. It treats quantization error as a stochastic perturbation, deriving a drift adjustment that preserves the marginal distribution of the generated output. This sampler-side correction improves metrics like FID without significant inference overhead.

Q-Drift corrects generation quality lost due to quantization noise accumulating during diffusion model sampling. It treats quantization error as a stochastic perturbation, deriving a drift adjustment that preserves the marginal distribution of the generated output. This sampler-side correction improves metrics like FID without significant inference overhead.

Open the full episode