Model Training
Inconsistent general-model output turned into repeatable, controlled visual languages.
- I did
- Dataset curation · training and tuning · prompting · art direction
- Built on
- Krea · ComfyUI · LoRA
- Proof
- Releases made with the trained systems surpassed 4M views
A set of custom-trained visual systems for producing consistent art direction across large bodies of generated work. The difficult part was making a general model repeat a specific visual language without flattening its range.
I curated and reduced the datasets, trained the LoRAs, tuned Krea 2 raw and turbo checkpoints in ComfyUI, shaped the prompts, and art-directed the releases. Krea and ComfyUI provided the model and training infrastructure; the resulting systems powered releases that surpassed 4M views on X.
