Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation

Diffusion distillation is a widely used technique to reduce the sampling cost of diffusion models, yet it often requires extensive training, and the student performance tends to be degraded. Recent studies show that incorporating a GAN objective may alleviate these issues, yet the underlying mechanism remains unclear. In this work, we first identify a key limitation of distillation: mismatched step sizes and parameter numbers between the teacher and the student model lead them to converge to different local minima, rendering direct imitation suboptimal. We further demonstrate that a standalone GAN objective, without relying a distillation loss, overcomes this limitation and is sufficient to convert diffusion models into efficient one-step generators. Based on this finding, we propose that diffusion training may be viewed as a form of generative pre-training, equipping models with capabilities that can be unlocked through lightweight GAN fine-tuning. Supporting this view, we create a one-step generation model by fine-tuning a pre-trained model with 85% of parameters frozen, achieving strong performance with only 0.2M images and near-SOTA results with 5M images. We further present a frequency-domain analysis that may explain the one-step generative capability gained in diffusion training. Overall, our work provides a new perspective for diffusion training, highlighting its role as a powerful generative pre-training process, which can be the basis for building efficient one-step generation models.

Distributed Rule Vectors is A Key Mechanism in Large Language Models' In-Context Learning

Large Language Models (LLMs) have demonstrated remarkable abilities, one of the most important being In-Context Learning (ICL). With ICL, LLMs can derive the underlying rule from a few demonstrations and provide answers that comply with the rule. Previous work hypothesized that the network creates a "task vector" in specific positions during ICL. Patching the "task vector" allows LLMs to achieve zero-shot performance similar to few-shot learning. However, we discover that such "task vectors" do not exist in tasks where the rule has to be defined through multiple demonstrations. Instead, the rule information provided by each demonstration is first transmitted to its answer position and forms its own rule vector. Importantly, all the rule vectors contribute to the output in a distributed manner. We further show that the rule vectors encode a high-level abstraction of rules extracted from the demonstrations. These results are further validated in a series of tasks that rely on rules dependent on multiple demonstrations. Our study provides novel insights into the mechanism underlying ICL in LLMs, demonstrating how ICL may be achieved through an information aggregation mechanism.

Diffusion Models Are Innate One-Step Generators

Diffusion Models (DMs) have achieved great success in image generation and other fields. By fine sampling through the trajectory defined by the SDE/ODE solver based on a well-trained score model, DMs can generate remarkable high-quality results. However, this precise sampling often requires multiple steps and is computationally demanding. To address this problem, instance-based distillation methods have been proposed to distill a one-step generator from a DM by having a simpler student model mimic a more complex teacher model. Yet, our research reveals an inherent limitations in these methods: the teacher model, with more steps and more parameters, occupies different local minima compared to the student model, leading to suboptimal performance when the student model attempts to replicate the teacher. To avoid this problem, we introduce a novel distributional distillation method, which uses an exclusive distributional loss. This method exceeds state-of-the-art (SOTA) results while requiring significantly fewer training images. Additionally, we show that DMs' layers are activated differently at different time steps, leading to an inherent capability to generate images in a single step. Freezing most of the convolutional layers in a DM during distributional distillation leads to further performance improvements. Our method achieves the SOTA results on CIFAR-10 (FID 1.54), AFHQv2 64x64 (FID 1.23), FFHQ 64x64 (FID 0.85) and ImageNet 64x64 (FID 1.16) with great efficiency. Most of those results are obtained with only 5 million training images within 6 hours on 8 A100 GPUs. This breakthrough not only enhances the understanding of efficient image generation models but also offers a scalable framework for advancing the state of the art in various applications.