Geração e Aperfeiçoamento de Logotipos via Ajuste Fino no Stable Diffusion XL

Abstract

This dissertation assesses fine-tuning techniques applied to Stable Diffusion XL (SDXL)—the high-resolution variant of the latent-diffusion model—with special emphasis on the lightweight approach offered by LoRA (Low-Rank Adaptation), which inserts low-dimensional matrices and therefore adds far fewer parameters. Six dedicated data sets were prepared: typographic structure (wordmark), graphic symbol (iconic), and three style filters (minimalistic, vintage, cartoon). Each set was trained for 10 epochs (an epoch is one complete pass through the entire training set), using different learning rates for the text modulator (CLIP-Text encoder) and for the image-generating U-Net.

The experimental phase produced 8,640 samples, systematically combining the Classifier-Free Guidance scale (CFG), the number of denoising steps (Steps), and the sampling algorithm (sampler). Performance was measured with CLIP similarity (semantic adherence), OCR accuracy (text legibility), and human evaluation. Results show that LoRA fine-tuning steers diffusion toward a narrower, more coherent solution space, reducing visual noise and unwanted variation. Text legibility rose from 37 % in the base model to 88 % after adjustment, and exceeded 92 % when an optional text-correction post-process (fix-text) was applied. Qualitative analysis confirms that the procedure preserves the intended style, improves brand-name accuracy, and allows a reduction in the total number of sampling steps. The study concludes that LoRA—coupled with careful data curation and fine hyper-parameter tuning—is the most effective alternative for specializing SDXL in automatic logo generation.