Beyond Fine-Tuning: A Systematic Study of Sampling Techniques in Personalized Image Generation
Keywords: Generative Model, Diffusion Model, Subject-Driven Generation
TL;DR: We propose a systematic and comprehensive analysis of different sampling techniques for personalized image generation and establish simple and strong baseline that outperforms or shows comparable results with existing personalization methods.
Abstract: Personalized text-to-image generation focuses on creating customized images based on user-defined concepts and text descriptions. A good balance between learned concept fidelity and its ability to be generated in different contexts is a major challenge in this task. Modern personalization techniques often strive to find this balance through diverse fine-tuning parameterizations and enhanced sampling methods that integrate superclass trajectories into the backward diffusion process. Improved sampling methods present a cost-effective, training-free way to enhance already fine-tuned models. However, outside of fine-tuning approaches, there is no systematic analysis of sampling methods in the personalised generation literature. Most sampling techniques are introduced alongside fixed fine-tuning parameterizations, which makes it difficult to identify the impact of sampling on the generation outcomes and whether it can be applied with other fine-tuning strategies. Moreover, they don't compare with the naive sampling approaches, so the intuition of how the superclass trajectory affects the sampling process remains underexplored. In this work, we propose a systematic and comprehensive analysis of personalized generation sampling strategies beyond the fine-tuning methods. We explore various combinations of concept and superclass trajectories, developing a deep understanding of how superclass influence generation outputs. Based on these results, we demonstrate that even a weighted mix of the concept and superclass trajectory can establish a strong baseline that enhances the adaptability of concepts across different contexts and can be effectively transferred to any training strategy, including various fine-tuning parameterizations, text embedding optimization, and hypernetworks. We analyze all methods through the lens of the trade-off between concept fidelity, editability, and computational efficiency, ultimately providing a framework to determine which sampling method is most suitable for specific scenarios.
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 11234
Loading