Angle Domain Guidance: Latent Diffusion Requires Rotation Rather Than Extrapolation

Cheng Jin, Zhenyu Xiao, Chutao Liu, Yuantao Gu
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:28187-28212, 2025.

Abstract

Classifier-free guidance (CFG) has emerged as a pivotal advancement in text-to-image latent diffusion models, establishing itself as a cornerstone technique for achieving high-quality image synthesis. However, under high guidance weights, where text-image alignment is significantly enhanced, CFG also leads to pronounced color distortions in the generated images. We identify that these distortions stem from the amplification of sample norms in the latent space. We present a theoretical framework that elucidates the mechanisms of norm amplification and anomalous diffusion phenomena induced by classifier-free guidance. Leveraging our theoretical insights and the latent space structure, we propose an Angle Domain Guidance (ADG) algorithm. ADG constrains magnitude variations while optimizing angular alignment, thereby mitigating color distortions while preserving the enhanced text-image alignment achieved at higher guidance weights. Experimental results demonstrate that ADG significantly outperforms existing methods, generating images that not only maintain superior text alignment but also exhibit improved color fidelity and better alignment with human perceptual preferences.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-jin25j, title = {Angle Domain Guidance: Latent Diffusion Requires Rotation Rather Than Extrapolation}, author = {Jin, Cheng and Xiao, Zhenyu and Liu, Chutao and Gu, Yuantao}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {28187--28212}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/jin25j/jin25j.pdf}, url = {https://proceedings.mlr.press/v267/jin25j.html}, abstract = {Classifier-free guidance (CFG) has emerged as a pivotal advancement in text-to-image latent diffusion models, establishing itself as a cornerstone technique for achieving high-quality image synthesis. However, under high guidance weights, where text-image alignment is significantly enhanced, CFG also leads to pronounced color distortions in the generated images. We identify that these distortions stem from the amplification of sample norms in the latent space. We present a theoretical framework that elucidates the mechanisms of norm amplification and anomalous diffusion phenomena induced by classifier-free guidance. Leveraging our theoretical insights and the latent space structure, we propose an Angle Domain Guidance (ADG) algorithm. ADG constrains magnitude variations while optimizing angular alignment, thereby mitigating color distortions while preserving the enhanced text-image alignment achieved at higher guidance weights. Experimental results demonstrate that ADG significantly outperforms existing methods, generating images that not only maintain superior text alignment but also exhibit improved color fidelity and better alignment with human perceptual preferences.} }
Endnote
%0 Conference Paper %T Angle Domain Guidance: Latent Diffusion Requires Rotation Rather Than Extrapolation %A Cheng Jin %A Zhenyu Xiao %A Chutao Liu %A Yuantao Gu %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-jin25j %I PMLR %P 28187--28212 %U https://proceedings.mlr.press/v267/jin25j.html %V 267 %X Classifier-free guidance (CFG) has emerged as a pivotal advancement in text-to-image latent diffusion models, establishing itself as a cornerstone technique for achieving high-quality image synthesis. However, under high guidance weights, where text-image alignment is significantly enhanced, CFG also leads to pronounced color distortions in the generated images. We identify that these distortions stem from the amplification of sample norms in the latent space. We present a theoretical framework that elucidates the mechanisms of norm amplification and anomalous diffusion phenomena induced by classifier-free guidance. Leveraging our theoretical insights and the latent space structure, we propose an Angle Domain Guidance (ADG) algorithm. ADG constrains magnitude variations while optimizing angular alignment, thereby mitigating color distortions while preserving the enhanced text-image alignment achieved at higher guidance weights. Experimental results demonstrate that ADG significantly outperforms existing methods, generating images that not only maintain superior text alignment but also exhibit improved color fidelity and better alignment with human perceptual preferences.
APA
Jin, C., Xiao, Z., Liu, C. & Gu, Y.. (2025). Angle Domain Guidance: Latent Diffusion Requires Rotation Rather Than Extrapolation. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:28187-28212 Available from https://proceedings.mlr.press/v267/jin25j.html.

Related Material