# Classifier-free Guidance with Adaptive Scaling

arXiv: https://arxiv.org/abs/2502.10574 · HTML: https://mlsanity.com/paper/2502.10574/

## Key facts

- Cited by 22 papers in the map (data through 14 Oct 2026).
- Placed in Diffusion foundations and sampling, one of 14 directions in Diffusion models.
- First posted on arXiv 14 Feb 2025.

## Placed in

- Diffusion models / [Diffusion foundations and sampling](https://mlsanity.com/diffusion/diffusion-foundations-and-sampling/)

## Built on by (most cited first, 22 of 22)

- [Stage-wise Dynamics of Classifier-Free Guidance in Diffusion Models](https://mlsanity.com/paper/2509.22007/) (2025)
- [Learn to Guide Your Diffusion Model](https://mlsanity.com/paper/2510.00815/) (2025)
- [UnGuide: Learning to Forget with LoRA-Guided Diffusion Models](https://mlsanity.com/paper/2508.05755/) (2025)
- [C2FG: Control Classifier-Free Guidance via Score Discrepancy Analysis](https://mlsanity.com/paper/2603.08155/) (2026)
- [Emergence of Distortions in High-Dimensional Guided Diffusion Models](https://mlsanity.com/paper/2602.00716/) (2026)
- [MMD Guidance: Training-Free Distribution Adaptation for Diffusion Models via Maximum Mean Discrepancy Guidance](https://mlsanity.com/paper/2601.08379/) (2026)
- [Memory Self-Regeneration: Uncovering Hidden Knowledge in Unlearned Models](https://mlsanity.com/paper/2510.03263/) (2025)
- [Analytic Distribution of Classifier-Free Guidance for Schedule Design](https://mlsanity.com/paper/2607.19725/) (2026)
- [RSTR: Reducing SpatioTemporal Redundancy in Diffusion Transformers](https://mlsanity.com/paper/2512.14096/) (2025)
- [DIAMOND: Directed Inference for Artifact Mitigation in Flow Matching Models](https://mlsanity.com/paper/2602.00883/) (2026)
- [Improving Diffusion Generalization with Weak-to-Strong Segmented Guidance](https://mlsanity.com/paper/2603.20584/) (2026)
- [Undress to Redress: A Training-Free Framework for Virtual Try-On](https://mlsanity.com/paper/2508.07680/) (2025)
- [Cross-modal Consistency Guidance for Robust Emotion Control in Auto-Regressive TTS Models](https://mlsanity.com/paper/2510.13293/) (2025)
- [QSilk: Micrograin Stabilization and Adaptive Quantile Clipping for Detail-Friendly Latent Diffusion](https://mlsanity.com/paper/2510.15761/) (2025)
- [ReLAPSe: Reinforcement-Learning-trained Adversarial Prompt Search for Erased concepts in unlearned diffusion models](https://mlsanity.com/paper/2602.00350/) (2026)
- [Guidance Is Not a Hyperparameter: Learning Dynamic Control in Diffusion Language Models](https://mlsanity.com/paper/2605.07701/) (2026)
- [Information-Theoretic Classifier-Free Guidance with Adaptive Schedule Optimization](https://mlsanity.com/paper/2606.24025/) (2026)
- [Adversarial Learning of Classifier-Free Guidance Schedules](https://mlsanity.com/paper/2608.14038/) (2026)
- [DCGC: Draft-Conditioned Global Correction for Complex Reasoning with Masked Diffusion Models](https://mlsanity.com/paper/2608.25428/) (2026)
- [Correcting Guided Diffusion Trajectories with Spectral Alignment](https://mlsanity.com/paper/2610.02753/) (2026)
- [Getting Your Guidance Weights Right in diffusion and flow-matching posterior sampling](https://mlsanity.com/paper/2610.03503/) (2026)
- [Opening the Black Box of Classifier-Free Guidance via Information Bottleneck](https://mlsanity.com/paper/2610.13369/) (2026)

## Abstract (arXiv)

Classifier-free guidance (CFG) is an essential mechanism in contemporary text-driven diffusion models. In practice, in controlling the impact of guidance we can see the trade-off between the quality of the generated images and correspondence to the prompt. When we use strong guidance, generated images fit the conditioned text perfectly but at the cost of their quality. Dually, we can use small guidance to generate high-quality results, but the generated images do not suit our prompt. In this paper, we present β-CFG (β-adaptive scaling in Classifier-Free Guidance), which controls the impact of guidance during generation to solve the above trade-off. First, β-CFG stabilizes the effects of guiding by gradient-based adaptive normalization. Second, β-CFG uses the family of single-modal (β-distribution), time-dependent curves to dynamically adapt the trade-off between prompt matching and the quality of samples during the diffusion denoising process. Our model obtained better FID scores, maintaining the text-to-image CLIP similarity scores at a level similar to that of the reference CFG.