model training

A style is a lock. Same palette, same grain, same register across thousands of frames, and the model stops guessing and starts repeating. Two tools got me there: Krea and ComfyUI.

krea

I trained custom styles on Krea 2 large and medium and used them as art direction tools. Pick the style, run it across a whole release, and every frame sits in the same universe.

comfyui

ComfyUI got me the control. The same datasets moved down to the checkpoint level, fine-tuning Krea 2 raw and turbo until the output was boring in the good way. Reproducible.

881K impressions. The fake fast-food post that started the run.

The release post cleared 1.1M views. The follow-up got 1M. Nine more posts followed: ps1 backrooms 358K, 2000s office 291K, summer 282K, optimistic backrooms 265K, fourth of July 188K, coffee 153K, character studies 107K. Past 4M in total.

Most of the work was deleting. Datasets get cut in half, prompts get discarded, two of the styles never shipped. Consistency is the boring advantage, and it is what you are paying for: anyone can pick the style up and get the same result.

The styles are public: krea 2 large and krea 2 medium.