Learning from human feedback has been shown to improve text-to-image models. These techniques first learn a reward function that captures what humans care about in the task and then improve the models based on the learned reward function. Even though relatively simple approaches (e.g., rejection sampling based on reward scores) have been investigated, fine-tuning text-to-image models with the reward function remains challenging. In this work, we propose using online reinforcement learning (RL) to fine-tune text-to-image models. We focus on diffusion models, defining the fine-tuning task as an RL problem, and updating the pre-trained text-to-image diffusion models using policy gradient to maximize the feedbacktrained reward. Our approach, coined DPOK, integrates policy optimization with KL regularization. We conduct an analysis of KL regularization for both RL fine-tuning and supervised fine-tuning. In our experiments, we show that DPOK is generally superior to supervised fine-tuning with respect to both image-text alignment and image quality.

Fine-tuning is arguably the most straightforward way to tailor a pre-trained model (e.g., a foundation model) to downstream applications, but it also comes with the risk of losing valuable knowledge the model had learned in pre-training. For example, fine-tuning a pre-trained classifier capable of recognizing a large number of classes to master a subset of classes at hand is shown to drastically degrade the model's accuracy in the other classes it had previously learned. As such, it is hard to further use the fine-tuned model when it encounters classes beyond the fine-tuning data. In this paper, we systematically dissect the issue, aiming to answer the fundamental question, What has been damaged in the fine-tuned model? To our surprise, we find that the fine-tuned model neither forgets the relationship among the other classes nor degrades the features to recognize these classes. Instead, the fine-tuned model often produces more discriminative features for these other classes, even if they were missing during fine-tuning! What really hurts the accuracy is the discrepant logit scales between the fine-tuning classes and the other classes, implying that a simple post-processing calibration would bring back the pre-trained model's capability and at the same time unveil the feature improvement over all classes. We conduct an extensive empirical study to demonstrate the robustness of our findings and provide preliminary explanations underlying them, suggesting new directions for future theoretical analysis.

We study the problem of transfer learning and fine-tuning in linear models for both regression and binary classification. In particular, we consider the use of stochastic gradient descent (SGD) on a linear model initialized with pretrained weights and using a small training data set from the target distribution. In the asymptotic regime of large models, we provide an exact and rigorous analysis and relate the generalization errors (in regression) and classification errors (in binary classification) for the pretrained and fine-tuned models. In particular, we give conditions under which the fine-tuned model outperforms the pretrained one. An important aspect of our work is that all the results are universal, in the sense that they depend only on the first and second order statistics of the target distribution. They thus extend well beyond the standard Gaussian assumptions commonly made in the literature. Furthermore, our universality results extend beyond standard SGD training to the test error of a classification task trained using ridge regression.

Text-to-image (T2I) diffusion models, when fine-tuned on a few personal images, can generate visuals with a high degree of consistency. However, such fine-tuned models are not robust; they often fail to compose with concepts of pretrained model or other fine-tuned models. To address this, we propose a novel fine-tuning objective, dubbed Direct Consistency Optimization, which controls the deviation between fine-tuning and pretrained models to retain the pretrained knowledge during fine-tuning. Through extensive experiments on subject and style customization, we demonstrate that our method positions itself on a superior Pareto frontier between subject (or style) consistency and image-text alignment over all previous baselines; it not only outperforms regular fine-tuning objective in image-text alignment, but also shows higher fidelity to the reference images than the method that fine-tunes with additional prior dataset. More importantly, the models fine-tuned with our method can be merged without interference, allowing us to generate custom subjects in a custom style by composing separately customized subject and style models. Notably, we show that our approach achieves better prompt fidelity and subject fidelity than those post-optimized for merging regular fine-tuned models.

When fine-tuning zero-shot models like CLIP, our desideratum is for the fine-tuned model to excel in both in-distribution (ID) and out-of-distribution (OOD). Recently, ensemble-based models (ESM) have been shown to offer significant robustness improvement, while preserving high ID accuracy. However, our study finds that ESMs do not solve the ID-OOD trade-offs: they achieve peak performance for ID and OOD accuracy at different mixing coefficients. When optimized for OOD accuracy, the ensemble model exhibits a noticeable decline in ID accuracy, and vice versa. In contrast, we propose a sample-wise ensembling technique that can simultaneously attain the best ID and OOD accuracy without the trade-offs.

Pre-trained foundation models, particularly large language models, have achieved remarkable success and led to massive fine-tuned variants. These models are commonly fine-tuned locally and then uploaded by users to cloud platforms such as HuggingFace for secure storage. However, the huge model number and their billion-level parameters impose heavy storage overhead for cloud with limited resources. Our empirical and theoretical analysis reveals that most fine-tuned models in cloud have a small difference (delta) from their pre-trained models. To this end, we propose a novel lossless compression scheme FM-Delta specifically for storing massive fine-tuned models in cloud.

However, the optimization of tasks performance often comes at the cost of generalizability in fine-tuned models. To address this issue, we theoretically connect smaller weight gradient norms during training and larger datasets to the improvements in model generalization. Motivated by this connection, we propose reducing gradient norms for enhanced generalization and aligning fine-tuned model with the pre-trained counterpart to retain knowledge from large-scale pre-training data. Yet, naive alignment does not guarantee gradient reduction and can potentially cause gradient explosion, complicating efforts to manage gradients. To address such an issue, we propose PACE, marrying generalization of PArameter-efficient fine-tuning with Consistency rEgularization. We perturb features learned from the adapter with the multiplicative noise and ensure the fine-tuned model remains consistent for same sample under different perturbations. Theoretical analysis shows that PACE not only implicitly regularizes gradients for enhanced generalization, but also implicitly aligns the fine-tuned and pre-trained models to retain knowledge. Experimental evidence supports our theories.

Large Language Models (LLMs) are typically trained in two phases: pre-training on large internet-scale datasets, and fine-tuning for downstream tasks. Given the higher computational demand of pre-training, it is intuitive to assume that fine-tuning adds less new information to the model, and is thus more compressible. We explore this assumption by decomposing the weights of fine-tuned models into their pre-trained components and an additional delta. We introduce a simple method, BitDelta, which successfully quantizes this delta down to 1 bit without compromising performance. This interesting finding not only highlights the potential redundancy of information added during fine-tuning, but also has significant implications for the multi-tenant serving and multi-tenant storage of fine-tuned models. By enabling the use of a single high-precision base model accompanied by multiple 1-bit deltas, BitDelta dramatically reduces GPU memory requirements by more than 10x, thus reducing per-user generation latency by more than 10x in multi-tenant settings. We validate BitDelta through experiments across Llama-2, Mistral and MPT model families, and on models up to 70B parameters, showcasing minimal performance degradation in all tested settings.