Open Access

Daniel Bittner

• This thesis investigates the influence of inference parameters on the visual quality of images generated by state‐of‐the‐art diffusion‐based models, with a particular focus on applications in game asset production. Motivated by the increasing prominence of generative AI in creative industries and the need for efficient, high‐quality 2D asset creation, this study addresses a critical gap in theThis thesis investigates the influence of inference parameters on the visual quality of images generated by state‐of‐the‐art diffusion‐based models, with a particular focus on applications in game asset production. Motivated by the increasing prominence of generative AI in creative industries and the need for efficient, high‐quality 2D asset creation, this study addresses a critical gap in the literature, which has predominantly concentrated on prompt optimization and text–image alignment. Two models, Stable Diffusion 3 Medium and Flux.1, were employed to systematically explore how variations in CFG scale, denoise strength, noise seed, and sampler–scheduler pairings affect both structural fidelity and perceptual quality. Multiple batches of images were generated under controlled parameter adjustments and subsequently evaluated using a comprehensive suite of image quality assessment metrics—including SSIM, MS‐SSIM, LPIPS, Laplacian variance, SIFT keypoints, and Earth Mover’s Distance (EMD) in both frequency and Lab domains. The results reveal that the CFG scale exerts a non‐linear effect on image quality, with mid‐range settings yielding optimal structural and perceptual similarity, while excessively low or high values lead to fragmentation or homogenization of details. Adjustments in denoise strength demonstrated a trade‐off between noise reduction and the preservation of fine image details, as excessive denoising improved clarity at the expense of textural nuance. Moreover, variations in the noise seed parameter induced significant stochastic variability in the outputs, and the selection of sampler–scheduler pairs was found to cause abrupt transitions in visual characteristics, underscoring their critical role in the generative process. These findings have important implications for the deployment of generative AI in practical settings, suggesting that fine‐tuning inference parameters is essential to balance creative flexibility with production consistency.…