AI Lab

Custom-trained style models, generation pipelines, and the tools I build to run them. Every model is trained on curated datasets and validated before production use.

• Case Study

  Character → Splash: the pipeline

  How an original character design becomes a key visual with its identity intact.

• Model Card

  Flat Casual Game Style — LoRA

  Style LoRA trained end-to-end: 1,167 images reviewed → 58 curated → epoch-swept, production-validated.

• Model Card

  Cinematic Painterly Realism — LoRA

  Low-saturation painterly realism for splash art and character concepts. 85-image dataset.

• Model Card

  Painterly Ornate Fantasy — LoRA

  Ornate semi-realistic fantasy character style. 150-image dataset, light-touch epoch selected.

• Model Card

  Chibi Game Style — LoRA

  Clean chibi game-art style. 60-image dataset, dual-epoch operation (stable / cute).

• Tool

  Atelier — generation workbench

  Self-built web app for batch review, comparison, and curation — 1,100+ batches, 4,100+ images and counting.

• Tool

  Datacurator — dataset pipeline

  Self-built tool for crawling, refining, captioning and exporting LoRA training datasets.