Research map / Image generation / Fast diffusion sampling

StepbaQ: Stepping backward as Correction for Quantized Diffusion Models

NeurIPS 2024 · Image generation › Fast diffusion sampling · approach: Fast diffusion ODE solvers · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes StepbaQ, a plug-and-play calibration method that treats quantization-induced latent distribution shift as a backward "stepback" in denoising and corrects the sampling steps and latents using error statistics from a 128-sample calibration set.

Problem

Quantized diffusion models accumulate quantization error across denoising steps, distorting the sampling trajectory and degrading image quality, and existing fixes are limited to stochastic samplers or only partially address the issue.

Approach

Interprets the latent distribution shift caused by quantization error (assumed Gaussian) as a temporal shift ("stepback") and finds a corrected timestep whose alpha matches the inflated noise level; it then adjusts the time embedding, latent (mean subtraction and rescaling), and step size, and corrects only when accumulated error is large enough. Applies to both deterministic and stochastic sample

Key results

• On SD v1.5 under W8A8, StepbaQ improves FID by 4.35 on MS-COCO and 7.30 on SDprompts over naive PTQ, while PTQD gives only 0.33 and 0.16.
• On SDXL-Turbo under W4A8, StepbaQ reduces FID from 44.76 to 23.92 on MS-COCO and from 45.45 to 28.14 on SDprompts.
• On SD v1.4 under W8A8 with the Q-Diffusion codebase, StepbaQ reaches FID 8.23 (MS-COCO) and 8.81 (SDprompts), beating Q-Diffusion, PTQD, and TFMQ-DM.

Measured effects reported in the paper

Surprising findings

• PTQD, a prior error-correction method, often underperforms the Q-Diffusion baseline on SD v1.4 because its correlated-noise elimination relies on weakly correlated estimates.
• Temporal information alignment does not help SDXL-Turbo, hypothesized to be because its training used only four fixed timesteps.

Limitations

• Does not apply to single-step generation such as 1-step SDXL-Turbo.
• Quantization error distribution has positive kurtosis (fat tails) rather than being exactly Gaussian, which the method's assumption ignores.
• No error bars or statistical significance reported.

Experimental setup

Datasets: MS-COCO Stable-Diffusion-Prompts (SDprompts)

Benchmarks: MS-COCO SDprompts

Metrics: FID CLIP score

Baselines: Naive PTQ PTQD Q-Diffusion TFMQ-DM

Builds on: PTQD DDIM Q-Diffusion TFMQ-DM

Closest papers by meaning

• AccuQuant: Simulating Multiple Denoising Steps for Quantizing Diffusion Models (NeurIPS 2025 · similarity 0.91)
• MPQ-DM: Mixed Precision Quantization for Extremely Low Bit Diffusion Models (AAAI 2025 · similarity 0.89)
• TCAQ-DM: Timestep-Channel Adaptive Quantization for Diffusion Models (AAAI 2025 · similarity 0.88)
• D^2-DPM: Dual Denoising for Quantized Diffusion Probabilistic Models (AAAI 2025 · similarity 0.88)
• Pioneering 4-Bit FP Quantization for Diffusion Models: Mixup-Sign Quantization and Timestep-Aware Fine-Tuning (CVPR 2025 · similarity 0.88)
• Qua2SeDiMo: Quantifiable Quantization Sensitivity of Diffusion Models (AAAI 2025 · similarity 0.88)