Recently, Diffusion Model has made significant progress in the field of image generation, bringing unprecedented development opportunities to image generation and video generation tasks. Despite the impressive results, the multi-step iterative denoising properties inherent in the inference process of diffusion models result in high computational costs. Recently, a series of diffusion model distillation algorithms have emerged to accelerate the inference process of diffusion models. These methods can be roughly divided into two categories: i) trajectory-preserving distillation; ii) trajectory reconstruction distillation. However, these two types of methods are limited by the limited effect ceiling or changes in the output domain.

In order to solve these problems, the ByteDance technical team proposed a trajectory segmentation consistency model called Hyper-SD. Hyper-SD's open source has also been recognized by Huggingface CEO Clem Delangue.

This model is a novel diffusion model distillation framework that combines the advantages of trajectory-preserving distillation and trajectory reconstruction distillation, while compressing the number of denoising steps. while maintaining near-lossless performance. Compared with existing diffusion model acceleration algorithms, this method achieves excellent acceleration results. Verified by extensive experiments and user reviews, Hyper-SD can achieve SOTA-level image generation performance in 1 to 8 steps of generation on both SDXL and SD1.5 architectures.

• Project homepage: https://hyper-sd.github.io/

• Paper link : https://arxiv.org/abs/2404.13686

• Huggingface Link: https://huggingface.co/ByteDance/Hyper-SD

• Single-step generated Demo link: https://huggingface.co/spaces/ByteDance/Hyper-SDXL-1Step-T2I

• Real-time drawing board Demo link: https://huggingface. co/spaces/ByteDance/Hyper-SD15-Scribble

##Introduction

Existing distillation methods for diffusion model acceleration can be roughly divided into two categories: trajectory-preserving distillation and trajectory reconstruction distillation. The trajectory-preserving distillation technique aims to maintain the original trajectory of the ordinary differential equation (ODE) corresponding to diffusion. The principle is to reduce inference steps by forcing the distilled model and the original model to produce similar outputs. However, it should be noted that although acceleration can be achieved, such methods may lead to a decrease in generation quality due to limited model capacity and inevitable errors in the training and fitting process. In contrast, trajectory reconstruction methods directly use the endpoints on the trajectory or real images as the main source of supervision, ignoring the intermediate steps of the trajectory, and can reduce the number of inference steps by reconstructing more effective trajectories and perform in a limited time. Explore the potential of your model within steps, freeing it from the constraints of the original trajectory. However, this often results in the output domain of the accelerated model being inconsistent with the original model, resulting in suboptimal results.

This paper proposes a trajectory segmentation consistency model (Hyper-SD for short) that combines the advantages of trajectory preservation and reconstruction strategies. Specifically, the algorithm first introduces trajectory segmentation consistency distillation to enforce consistency within each segment and gradually reduces the number of segments to achieve full-time consistency. This strategy solves the problem of suboptimal performance of consistent models due to insufficient model fitting capabilities and accumulation of inference errors. Subsequently, the algorithm uses human feedback learning (RLHF) to improve the model generation effect to make up for the loss of model generation effect during the acceleration process and make it better adapted to low-step reasoning. Finally, the algorithm uses fractional distillation to enhance one-step generation performance and achieves an idealized full-time-step consistent diffusion model through unified LORA, achieving excellent results in generation effects.

Method

1. Trajectory segmentation consistency distillation

Consistency Distillation (CD) [24] and Consistency Trajectory Model (CTM) [4] both aim to transform the diffusion model into a consistency model for the entire time step range [0, T] through one-shot distillation. However, these distillation models often fail to achieve optimality due to limitations in model fitting capabilities. Inspired by the soft consistency objective introduced in CTM, we refine the training process by dividing the entire time step range [0, T] into k segments and performing piecewise consistent model distillation step by step.

In the first stage, we set k=8 and use the original diffusion model to initialize and . The starting time step is uniformly randomly sampled from . We then sample the end time step , where is calculated as follows:

The training loss is calculated as follows:

Where is calculated by formula 3, represents the exponential moving average (EMA) of the student model.

Subsequently, we restore the model weights from the previous stage and continue training, gradually reducing k to [4,2,1]. It is worth noting that k=1 corresponds to the standard CTM training scheme. For the distance metric d, we employ a mixture of adversarial loss and mean squared error (MSE) loss. In experiments, we observed that the MSE loss is more effective when the predicted and target values ??are close (e.g., for k=8, 4), while the adversarial loss increases as the difference between the predicted and target values ??increases. becomes more precise (for example, for k=2, 1). Therefore, we dynamically increase the weight of the adversarial loss and decrease the weight of the MSE loss throughout the training phase. In addition, we also integrate a noise perturbation mechanism to enhance training stability. Take the two-stage Trajectory Segment Consensus Distillation (TSCD) process as an example. As shown in the figure below, we perform independent consistency distillation in the and time periods in the first stage, and then perform global consistency trajectory distillation based on the consistency distillation results of the previous two periods.

The complete algorithm process is as follows:

2. Human feedback learning

In addition to distillation, we further incorporate feedback learning to improve the performance of the accelerated diffusion model. Specifically, we improve the generation quality of accelerated models by leveraging human aesthetic preferences and feedback from existing visual perception models. For aesthetic feedback, we utilize the LAION aesthetic predictor and the aesthetic preference reward model provided in ImageReward to guide the model to generate more aesthetic images, as follows:

Where is the aesthetic reward model, including the aesthetic predictor of the LAION dataset and ImageReward model, c is the text prompt, is used together with the ReLU function as the hinge loss. In addition to feedback from aesthetic preferences, we note that existing visual perception models embedding rich prior knowledge about images can also serve as good feedback providers. Empirically, we find that instance segmentation models can guide the model to generate well-structured objects. Specifically, we first diffuse the noise on image in the latent space to , after which, similar to ImageReward, we perform iterative denoising until a specific time step and predict directly. We then leverage a perceptual instance segmentation model to evaluate the performance of structure generation by examining the difference between instance segmentation annotations for real images and instance segmentation predictions for denoised images, as follows:

Where is the instance segmentation model (such as SOLO). Instance segmentation models can more accurately capture the structural defects of generated images and provide more targeted feedback signals. It is worth noting that in addition to instance segmentation models, other perceptual models are also applicable. These perceptual models can serve as complementary feedback to subjective aesthetics, focusing more on objective generative quality. Therefore, the diffusion model we use to optimize the feedback signal can be defined as:

3. One-step generation of enhanced

due to consistency loss Inherent limitations, one-step generation within the consistency model framework are not ideal. As analyzed in CM, the consistent distillation model shows excellent accuracy in guiding the trajectory endpoint at position . Therefore, fractional distillation is a suitable and effective method to further improve the one-step generation effect of our TSCD model. Specifically, we advance further generation through an optimized distribution matching distillation (DMD) technique. DMD enhances the model's output by utilizing two different scoring functions: distribution from the teacher model and from the fake model. We combine mean squared error (MSE) loss with score-based distillation to improve training stability. In this process, the aforementioned human feedback learning techniques are also integrated to fine-tune our model to effectively generate images with high fidelity.

By integrating these strategies, our method can not only achieve excellent low-step inference results on both SD1.5 and SDXL (and without Classifier-Guidance), but also achieve an ideal global consistency model without targeting Each specific number of steps trains UNet or LoRA to implement a unified low-step reasoning model.

Experiment

On SD1.5 and SDXL and the current existing A quantitative comparison of acceleration algorithms shows that Hyper-SD is significantly better than the current state-of-the-art methods

In addition, Hyper-SD can use one model to implement various Different from low-step inference, the above quantitative indicators also show the effect of our method when using unified model inference.

The visualization of the acceleration effect on SD1.5 and SDXL intuitively demonstrates the superiority of Hyper-SD in accelerating diffusion model inference sex.

A large number of User-Study also shows the superiority of Hyper-SD compared to various existing acceleration algorithms.

The accelerated LoRA trained by Hyper-SD is well compatible with different styles of Vincent figure base models.

At the same time, Hyper-SD’s LoRA can also adapt to the existing ControlNet to achieve high-quality controllable image generation at a low number of steps.

Summary

The paper proposes Hyper-SD, a unified diffusion model acceleration framework that can significantly improve the generation ability of diffusion models in low-step situations. , realizing new SOTA performance based on SDXL and SD15. This method uses trajectory segmentation consistency distillation to enhance the trajectory preservation capability during the distillation process and achieve a generation effect close to the original model. Then, the potential of the model at extremely low step counts is improved by further leveraging human feedback learning and variational fractional distillation, resulting in more optimized and efficient model generation. The paper also open sourced the Lora plug-in for SDXL and SD15 from 1 to 8 steps inference, as well as a dedicated one-step SDXL model, aiming to further promote the development of the generative AI community.

The above is the detailed content of Accelerate diffusion model, generate SOTA-level images in the fastest 1 step, Byte Hyper-SD is open source. For more information, please follow other related articles on the PHP Chinese website!