GTA V Styled LoRA

Origin

My brother asked if I could replicate the GTA V art style in AI-generated images. That was the entire brief.

What I Built

Trained a GTA V style LoRA from scratch using SDXL and ComfyUI - data collection, captioning, hyperparameter tuning, evaluation.

The training process taught me more about diffusion model fine-tuning than any tutorial had. Specifically:

• How training data quality and variety affect style generalization
• How captioning strategy changes what the model learns to associate with the trigger token
• How learning rate and step count interact - overtrained models lose generalization, undertrained ones don't capture the style consistently

The output style is consistent enough to use in other experiments - which is how it ended up in the Mooood of Elon project.