Towards Highly Realistic Artistic Style Transfer via Stable Diffusion with Step-aware and Layer-aware Prompt

TL;DR

LSAST addresses the challenge of preserving content structure in diffusion-based artistic style transfer by introducing Step-aware and Layer-aware Prompts, enabling multi-scale conditioning across diffusion steps and U-Net layers. It learns style priors from artwork collections via a Step-aware and Layer-aware Prompt Inversion and leverages a Content Prompt from ControlNet during inference to improve structural fidelity. Empirical results show LSAST produces more realistic stylizations with fewer artifacts and better content preservation than state-of-the-art GAN- and diffusion-based methods, validated by FID, user preference, and deception metrics. The framework provides a scalable, controllable approach to artistic style transfer that can extend to other diffusion backbones and content prompts.

Abstract

Artistic style transfer aims to transfer the learned artistic style onto an arbitrary content image, generating artistic stylized images. Existing generative adversarial network-based methods fail to generate highly realistic stylized images and always introduce obvious artifacts and disharmonious patterns. Recently, large-scale pre-trained diffusion models opened up a new way for generating highly realistic artistic stylized images. However, diffusion model-based methods generally fail to preserve the content structure of input content images well, introducing some undesired content structure and style patterns. To address the above problems, we propose a novel pre-trained diffusion-based artistic style transfer method, called LSAST, which can generate highly realistic artistic stylized images while preserving the content structure of input content images well, without bringing obvious artifacts and disharmonious style patterns. Specifically, we introduce a Step-aware and Layer-aware Prompt Space, a set of learnable prompts, which can learn the style information from the collection of artworks and dynamically adjusts the input images' content structure and style pattern. To train our prompt space, we propose a novel inversion method, called Step-ware and Layer-aware Prompt Inversion, which allows the prompt space to learn the style information of the artworks collection. In addition, we inject a pre-trained conditional branch of ControlNet into our LSAST, which further improved our framework's ability to maintain content structure. Extensive experiments demonstrate that our proposed method can generate more highly realistic artistic stylized images than the state-of-the-art artistic style transfer methods.

Figures (5)

• Figure 1: Stylization examples with three different styles (i.e., Van Gogh, Monet, Gauguin). Compared to existing state-of-the-art large-scale pre-trained diffusion model-based methods (b-d) and generative adversarial network-based methods (e-h), our proposed method (a) generates highly realistic artistic stylized images and preserves the content structure of input content images well.
• Figure 2: The overview of our proposed framwork which consists of a training stage and an inference stage. In the training stage, the Step-aware and Layer-aware Prompt Spaces will learn and store the style information from the collection of artworks. In the inference stage, we utilize a pre-trained conditional branch of ControlNet, as content prompt, into our LSAST, which further improved our framework's ability to maintain content structure.
• Figure 3: Differences between (a) ArtBank. (b) ProSpect. (c) LSAST (Ours). Unlike existing methods that use only a global/step-aware prompt to condition the whole diffusion process, LSAST utilizes a step-aware and layer-aware prompt to dynamically adjusts the input images' content structure and style pattern.
• Figure 4: Qualitative comparisons with other state-of-the-art artistic style methods. The first column shows the input content image. The (b-d) column shows the stylized image from large-scale pre-trained diffusion model-based methods, and the (e-h) column presents the stylized images generated by generative adversarial network-based methods.
• Figure 5: Ablation studies of our LSAST. Please zoom-in for better comparison.