Research map / Image generation / Fast diffusion sampling

Rectified Diffusion Guidance for Conditional Generation

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

Shows that summing-to-one classifier-free guidance breaks the diffusion theory (expectation shift) and proposes ReCFG, a post-hoc relaxation of guidance coefficients with a pixel-wise closed-form lookup-table correction for γ0 that is applied without retraining.

Problem

Classifier-free guidance combines conditional and unconditional scores with coefficients summing to one, which is not a reciprocal of the diffusion process and causes expectation shift of the generated distribution, harming conditional fidelity especially at large guidance strengths.

Approach

Relax the CFG coefficients to independent γ1 (conditional) and γ0 (unconditional), and choose γ0 in closed form from the expectation ratio of the conditional and unconditional noise predictions, pre-computed by traversing data per condition, so that the expectation of the sampled distribution matches the ground truth.

Key results

• On ImageNet 512x512, ReCFG lowers FID for EDM2-L from 1.96 to 1.89 at 63 NFE with the same guidance strength.
• On CC12M with SD3 at 5 NFE and γ1=7.5, ReCFG raises CLIP-S from 0.209 to 0.229 and lowers FDDINOv2 from 1466.91 to 1323.49 versus CFG.
• On LDM ImageNet 256 at 10 NFE with γ1=3.0, ReCFG reduces FID from 10.13 to 8.26 relative to CFG.

Measured effects reported in the paper

Surprising findings

• Theory shows that with summing-to-one CFG at γ>1 the toy distribution's expectation is at least twice the ground truth.
• Expectation ratios vary strongly across pixels and show no general pattern, so scalar coefficients are argued to be insufficient.

Limitations

• Requires pre-computing a lookup table by traversing data per condition, which is impractical for open-vocabulary text conditions.
• Variance of the sampled distribution cannot be analytically controlled; variance constraints are set empirically.
• Theory for variance is only established on a 1-D toy Gaussian example.

Experimental setup

Datasets: ImageNet CC12M MS-COCO validation

Benchmarks: ImageNet 256x256 ImageNet 512x512 CC12M 512x512

Metrics: FID FDDINOv2 Improved Precision Improved Recall CLIP-S MPS

Baselines: CFG RescaleCFG

Builds on: classifier-free guidance DDIM EDM2 RescaleCFG

Closest papers by meaning

• Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.93)
• Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.93)
• No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.93)
• CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.92)
• TCFG: Tangential Damping Classifier-free Guidance (CVPR 2025 · similarity 0.92)
• Diffusion Models without Classifier-free Guidance (NeurIPS 2026 · similarity 0.91)