Diffusion-based valuable NFT generation

Multimedia Tools and Applications (2026) 85:557 https://doi.org/10.1007/s11042-026-21724-6 Diffusion-based valuable NFT generation Emir Ulurak1 · Beyza Kaya1 · Emre Sefer1 Received: 19 November 2024 / Revised: 29 January 2026 / Accepted: 1 June 2026 © The Author(s) 2026 Abstract Non-fungible tokens (NFTs) have revolutionized digital ownership, offering unique provenance and value to digital assets. Existing text-to-image models do not have the incentive mechanisms to generate statistically rare features, even when they optimize for visual fidelity. This paper introduces DiffNFTGen, a new generative framework that is the first to combine a customized RarityReward measure derived from a Vision Transformer (ViT) with reinforcement learning. The suggested method ensures fidelity to NFT styles while explicitly maximizing the generation of rare features by fine-tuning Stable Diffusion using Proximal Policy Optimization (PPO) and Kullback-Leibler (KL) divergence regularization. DiNFTGen achieves a 2.4x greater rarity score than baseline models while keeping competitive visual quality, according to quantitative evaluation utilizing Frédechet Inception Distance (FID) and Rarity Score. In order to examine the trade-off between fidelity and rarity, we also perform ablation studies regarding reward weighting. The model’s capacity to generalize NFT styles to new domains is confirmed by qualitative evaluations. The datasets, analysis code, and suggested approach are accessible on https://github.com /seferlab/diffnftgen. Keywords NFTs · Diffusion models · Reinforcement learning · Text to image generation Emir Ulurak and Beyza Kaya contributed equally to this work. Emre Sefer Emir Ulurak Beyza Kaya 1 Artificial Intelligence and Data Engineering Department, Ozyegin University, Orman Sokak, Cekmekoy 34794, Istanbul, Turkey 557 Page 2 of 23 Multimedia Tools and Applications (2026) 85:557 1 Introduction Non-fungible tokens (NFTs) are blockchain-based tokens that may represent a wide range of assets such as art pieces, digital content, video, media, etc. [1]. Creating Non-Fungible Token (NFT) images has gained tremendous popularity in recent years, because of their visual uniqueness, attractiveness, and richness of various gorgeous elements. NFT certificates represent irreversible and unique authenticity and ownership certificates. Even though NFTs can define ownership of different digital items, they have mainly been adapted to images and digital art. NFTs have been minted in terms of collections, which have been bought and sold with auctions [2, 3]. It is reported that the NFT market is expected to significantly grow at an annual rate of 35.0%, reaching $13.6 billion by 2027. From more of a financial perspective, NFTs can also be seen as an alternative investment class [4]. Although humans have generated many profitable collections, their systematic generation through artificial intelligence remains limited. Even though art-style NFT images are quite popular, designing such artistic-style images is a demanding task since art will generate something unique as well as appealing visually. Human artists should take into account a number of elements while designing, such as unique generation, rarity, and artistic and aesthetic appeal. These elements are required for NFTs to appear valuable in the competitive NFT market. Additionally, since there is an increasing demand for NFT images, the need for innovation and unique designs is increasing, which is important to catch the NFT market’s attention. In this case, NFT markets show the demand from potential NFT collectors and buyers. AI-based image generation has been extensively studied, where promising results have been obtained via Variational AutoEncoders [5, 6], generative adversarial networks (GANs) [7–9], and more recent diffusion-based models [10–15]. Among them, Stable Diffusion [16] has achieved the best performance where diffusion is performed in a latent space. Stable Diffusion not only has a lower computational cost, but its visual performance is remarkable. In this paper, we introduce DiffNFTGen, a novel generative deep reinforcement learningbased model, which is designed to create high-quality NFT images with a focus on rarity. Building upon the Stable Diffusion framework [16], DiffNFTGen combines reinforcement learning with a custom RarityReward metric created using the Vision Transformer (ViT) model [17] to obtain the rarity values of NFTs. DiffNFTGen employs Proximal Policy Optimization (PPO) [18] to address complexity of the reinforcement learning model. By limiting the scope of policy updates, PPO (PPO) [18] enables more gradual and stable learning. Additionally, the approach incorporates Kullback-Leibler (KL) divergence [19] regularization to ensure that the generated images adhere closely to the original NFT styles while also promoting their uniqueness and value. DiffNFTGen is designed to generate NFT images in response to user prompts, with a specific focus on integrating features that enhance the perceived value of these NFTs. In this case, our goal is to generate rare NFTs since the rarity of an NFT has been identified as a crucial determinant of its market value among the various attributes. We compare DiffNFTGen with other alternative diffusion models [20] which has limited capacity for our NFT generation’s comprehensive needs. Relative to the DPOK [21] framework, which combines reinforcement learning techniques with a diffusion model, 13 Multimedia Tools and Applications (2026) 85:557 Page 3 of 23 557 alternative diffusion-based approaches do not fully align with our dataset and the unique requirements of the rarity feature we aim to incorporate. Once we transform the rarity feature into a quantitative evaluation metric, we facilitate its integration into DiffNFTGen. As part of experimenting with various metrics to ensure that the generated images maintained high quality without significant fidelity loss, we found that the specificity of prompts plays a critical role in the quality of the generated images. Moreover, detailed prompts and real NFT references significantly enhanced the output, highlighting the importance of prompt engineering in our generative process. Overall, these advancements position DiffNFTGen as a significant contribution to the field of NFT generation, offering a model that meets market demands and adheres to high aesthetic standards. 1.1 Contributions To summarize, the main contributions of DiffNFTGen are as follows: ● Objective Definition of Generative Value: We propose a formalized definition of "value" for generated NFTs based on statistical rarity rather than speculative market pricing. By calculating attribute rarity weights derived from collection-level trait distributions, we establish a quantitative ground truth for uniqueness. Instead of depending only on subjective aesthetic quality, this enables the generating process to optimize for quantitative distinctiveness. ● (...truncated)