Research map / Image generation / Fast diffusion sampling

REG: Rectified Gradient Guidance for Conditional Diffusion Models

ICML 2025 · Image generation › Fast diffusion sampling · approach: Guidance for conditional diffusion · 0 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.

Reframes diffusion guidance as approximating an optimal joint-distribution scaling and proposes REG, a gradient-based correction to CFG/AutoG/CG that improves FID and IS/CLIP on ImageNet and COCO text-to-image.

Problem

Guidance in conditional diffusion models is practically implemented in a way that does not match its theoretical motivation: scaling marginal distributions is shown to be invalid, so the sampler does not target the intended scaled distribution.

Approach

Replace the invalid marginal-scaling objective with a valid joint-distribution scaling objective, derive the exact guidance noise predictor (requiring future foresight), interpret existing guidance as its no-foresight approximation, and add a Jacobian-based gradient correction term (REG) that better approximates the optimal guidance.

Key results

• On ImageNet 256x256 with DiT-XL/2, REG with vanilla CFG lowers FID from 2.21 to 2.04 and raises IS from 248.36 to 276.26.
• On COCO-2017 5K with SD-XL, REG with vanilla CFG improves FID from 23.73 to 23.46 and CLIP score from 31.38 to 31.51; with cosine CFG FID drops from 32.14 to 25.62.
• REG adds roughly 1.2x to 1.9x runtime and 1.2x to 2.8x peak memory over CFG on a single A40 GPU.

Measured effects reported in the paper

Surprising findings

• Cosine and linear CFG underperform vanilla CFG on COCO text-to-image, and SD-XL performs worse than SD-v1-4 on that benchmark.
• The guidance scheme implies the denoising start distribution should be modified, not kept as a standard Gaussian, which the authors leave as an open question.

Limitations

• REG requires a Jacobian-vector computation, adding memory and runtime overhead (up to 2.82x memory for SD-XL).
• The exact optimal guidance is intractable; REG uses diagonal and other approximations that are justified empirically.
• The modification of the denoising start distribution is not practically implemented.

Experimental setup

Datasets: ImageNet COCO-2017

Benchmarks: ImageNet 64x64 ImageNet 256x256 ImageNet 512x512 COCO-2017 5K

Metrics: FID Inception Score CLIP score runtime peak memory

Baselines: vanilla CFG cosine CFG linear CFG interval CFG AutoG classifier guidance

Builds on: classifier-free guidance classifier guidance auto-guidance DDPM interval guidance

Closest papers by meaning

• Feedback Guidance of Diffusion Models (NeurIPS 2025 · similarity 0.91)
• Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.91)
• Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.91)
• Towards a Golden Classifier-Free Guidance Path via Foresight Fixed Point Iterations (NeurIPS 2025 · similarity 0.91)
• Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models (NeurIPS 2024 · similarity 0.91)
• Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.90)