Unlocking Advanced Image Generation with Diffusion Fuzzy Systems

In recent years, the rapid development of artificial intelligence has advanced image generation technology significantly. Emerging generative models, particularly diffusion models, have not only consistently enhanced the quality of generated images but also broadened the application domains to such areas as artistic creation [1], virtual reality [2], and medical imaging [3]. Diffusion models are currently the leading approach in image generation. Traditional single-path models, like Denoising Diffusion Probabilistic Model (DDPM) [4], generate images by adding noise in a forward process and remove it step by step in a reverse process. To reduce the high sampling cost of DDPM, Denoising Diffusion Implicit Models (DDIM) [5] introduces a non-Markovian forward process for faster sampling, but the training and inference remain expensive. Latent Diffusion Models (LDM) [6] further reduce computation by performing diffusion in a compressed latent space. Classical diffusion models often produce lower-quality images due to limited guidance during diffusion. To address this limitation, Dhariwal et al. proposed Ablated Diffusion Model with Classifier Guidance (ADM-G) [7] to integrate a classifier into the reverse process. At each step, sampling was guided by gradients from the classifier to enable class-conditional generation without retraining the model. To realize conditional generation from images, text, or multimodal inputs, Liu et al. proposed Semantic Diffusion Guidance (SDG) [8] using content and style guidance via gradients. Classifier-guided diffusion, however, is limited by the need to train the classifier and diffusion model separately. This limitation is avoided by Classifier-Free Diffusion Guidance (CFDG) [9] which uses implicit guidance, and by Guided Language to Image Diffusion for Generation and Editing (GLIDE) [10] which combines classifier-free guidance with Contrastive Language-Image Pretraining (CLIP) [11] for large-scale text-to-image generation. Unconditional Contrastive Language-Image Pretraining model (unCLIP) [12] further maps text to images for salient content generation. Despite the success, these single-path models still struggle with images having highly heterogeneous features. Bar-Tal et al. proposed Multi-path Diffusion (MD) [13] to decompose images into subregions for parallel multi-path diffusion and fuse the outputs based on attention mechanism. Xue et al. proposed Regions Align with different text Phases in Attention Learning (RAPHAEL) [14], which further improved local details by dynamically routing heterogeneous diffusion paths based on text complexity. Residual Denoising Diffusion Model (RDDM) [15] enhances denoising robustness through residual learning. Despite these advances, multi-path models still have difficulty with globally inconsistent outputs when features vary greatly across images and focus mainly on local details. They also face challenges in inter-path coordination and higher computational cost. Although multi-path diffusion models can better capture different regions than single-path models, they are not effective for image collections with large feature differences. As shown in Fig. 1(a), multi-path diffusion models can generate coherent images for similar categories (e.g., landscapes), but when dealing with images of diverse categories (e.g., animals, humans, landscapes), as shown in Fig. 1(b), they only focus on local regions and fail to model global features and produce unrealistic results. Multi-path modeling also brings challenges in inter-path coordination and higher computational cost. Current multi-path diffusion models face three main challenges: they often capture local features only and miss global differences, struggle to coordinate multiple paths without causing mode collapse, and incur high computational costs that increase with the number of paths. To address the limitations of existing multi-path diffusion models, this paper proposes a novel approach called the Diffusion Fuzzy System (DFS), which integrates the architectures of traditional fuzzy systems and diffusion models. Essentially, DFS can be regarded as a fuzzy-rule-guided, latent-space multi-path diffusion modeling framework. In DFS, multiple diffusion paths in latent space are built using fuzzy rules, capturing uncertainty in image generation. The main contributions of this work are summarized as follows: i) We are the first to propose the integration of fuzzy systems with diffusion models to extend traditional fuzzy system framework to diffusion learning scenarios. Diffusion Fuzzy System is proposed to implement a fuzzy-rule-guided latent-space multi-path diffusion model. It captures the uncertainty and ambiguity inherent in generative tasks explicitly, which offers a novel perspective and methodology for uncertainty modeling in generative tasks. ii) A fuzzy rule-chain mechanism for diffusion learning is proposed to integrate cascaded fuzzy rules at each step to provide interpretable guidance. This extends the use of fuzzy rules in generative tasks and offers a new approach for guiding diffusion models. iii) A fuzzy-membership-based latent-space compression mechanism is introduced using adaptive encoder-decoder selection to map images into low-dimensional features. This reduces computational and memory costs while ensuring accurate reconstruction. This improves the practicality of multi-path diffusion models in resource-limited or real-time scenarios. The remainder of this paper is organized as follows. Section II introduces the fundamental concepts and principles of fuzzy systems and diffusion models. Section III presents the proposed Diffusion Fuzzy System (DFS). Section IV evaluates DFS with comprehensive experiments. Finally, Section V concludes the study and discusses future research directions.