Portfolio AI

AI training is the process of teaching a model to recognize patterns and generate results based on large sets of data — for example, images, text, or sounds. LoRA training (Low-Rank Adaptation) is a specialized and efficient method used to fine-tune large models like Stable Diffusion without retraining them from scratch. By training LoRAs on targeted image sets, it’s possible to create custom visual styles, characters, or product looks while keeping the base model lightweight and flexible. This approach allows for rapid experimentation and precise creative control in AI-assisted production. For this project, I chose the HiDream model for two main reasons: it’s open source and it captures the characteristics of its training data very well. Training a Hat LoRA was quite challenging. Since most people today don’t wear hats, the major AI models lack sufficient training data featuring them. This limitation forced me to get creative with my training datasets — though I won’t reveal all of my tricks. The results are not perfect yet, but they’re showing clear improvement. I’m currently experimenting with new models and will update this project as soon as I achieve better results.