February 15, 2026

Overview

Trained custom AI image models and LoRAs (Low-Rank Adaptations) for Apatero Studio and freelance projects. Focus on achieving specific visual styles, character consistency, and brand-aligned outputs.

Model Experience

• Flux: Fine-tuning for photorealistic and artistic generation with strong prompt adherence.
• Stable Diffusion (1.5 through XL): LoRA training across all major versions including Turbo and Lightning variants.
• Custom Datasets: Built and curated training datasets from scratch, including captioning, preprocessing, and quality filtering.

What I Train

• Style LoRAs: Models that replicate specific artistic styles or photography techniques.
• Character LoRAs: Consistent character representations across multiple generations.
• Product LoRAs: Brand-specific models for e-commerce and marketing content.
• Concept LoRAs: Teaching models new objects or scenarios not in their base training data.

Technical Setup

• Training Tools: Kohya Scripts, Hugging Face Diffusers, DreamBooth
• Infrastructure: GPU clusters, Docker, cloud compute
• Optimization: VRAM-efficient training techniques, hyperparameter tuning
• Quality Control: FID/CLIP score tracking, systematic A/B comparison of training runs

Results

• Reduced training iteration time by 60% through pipeline optimization.
• Trained models deployed in production on Apatero Studio, serving real users daily.
• Built reusable training pipelines that can be adapted for new model architectures as they release.