Research map / Image generation / Fast diffusion sampling

Feedback Guidance of Diffusion Models

NeurIPS 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.

Proposes FeedBack Guidance (FBG), a state- and time-dependent guidance scale derived from an additive error model and a posterior estimated by tracking the diffusion Markov chain, evaluated on EDM2-XS ImageNet512 and SD2 text-to-image.

Problem

Classifier-free guidance applies a constant scale at all noise levels and states, which can reduce diversity, induce memorization, and over-saturate samples even when the conditional model is already adequate.

Approach

Replace CFG's implicit multiplicative error assumption with an additive mixture of conditional and unconditional distributions, yielding a guidance scale lambda(x_t,t) computed from a posterior p(c|x_t) that is estimated online by tracking the likelihood ratio of conditional vs unconditional denoising steps along the sampling chain.

Key results

• On ImageNet512 with EDM2-XS (stochastic sampler), FBG_LIG reaches FD_DINOv2 87.9 vs CFG 100.2, and FBG_pure reaches FID 3.76 vs CFG 5.00.
• On MS-COCO with SDv2, FBG_CFG reaches FID 18.63 and FD_DINOv2 52.14 versus CFG 19.64 and 54.56.
• FBG applies the strongest average guidance to the hardest prompt difficulty levels in a 60-prompt set of four difficulty classes.

Measured effects reported in the paper

Surprising findings

• Late-stage guidance is found to be not only unnecessary but harmful, particularly for FD_DINOv2, contrary to prior work.

Limitations

• The additive mixture error model and prior are chosen for simplicity and may not approximate true model biases.
• Evaluation focuses on EDM2-XS; larger architectures and broader T2I prompt datasets are untested.
• Text-to-image evaluation is limited to MS-COCO, which has simple captions, and is shown qualitatively with a small 60-prompt set.

Experimental setup

Datasets: ImageNet512x512 MS-COCO

Benchmarks: ImageNet512x512 MS-COCO

Metrics: FID FD_DINOv2 Precision Recall Aesthetic Score

Baselines: CFG CFG++ Linear guidance weight scheduler LIG LinCFG

Builds on: Classifier-Free Guidance Limited Interval Guidance Dynamic Negative Guidance EDM2

Closest papers by meaning

• Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.92)
• Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.92)
• Diffusion Models without Classifier-free Guidance (NeurIPS 2026 · similarity 0.92)
• No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.91)
• REG: Rectified Gradient Guidance for Conditional Diffusion Models (ICML 2025 · similarity 0.91)
• Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.91)