About the lab

A lab for images made with AI.

The ai.doll lab is an open logbook. Every project starts from a question, and what I publish is the road to the answer: the batches, the attempts I threw away, the settings I used and what each one taught me.

The finished image is the way in. The process is what I actually want to share.

Line-drawn portrait of the author: glasses, dark tousled hair and a beard, smiling, in black and white.

Who's behind this

UX designer and design leader with years of experience shaping digital products, teams and systems, always driven by curiosity and by the need to understand how people interact with technology.

The ai.doll lab extends that mindset into a more personal and experimental space, where visual culture, character design and generative AI become tools to explore new creative processes, test ideas and build a point of view.

More than producing images, it is a way of learning through making, combining design thinking with artistic experimentation, and questioning how creativity, authorship and technology are evolving together.

How I work

  1. 01

    A question first

    Not «make something pretty», but something I can answer with images: how far can I change a character's world before she stops being herself?

  2. 02

    Batches, not lucky shots

    One experiment can be twelve batches and two hundred images to keep ten. The ones that failed are part of the answer, so they stay.

  3. 03

    Everything on one contact sheet

    Each phase becomes a single sheet with every attempt in a grid, each with its code and its verdict. You see what I saw, in the order I saw it.

  4. 04

    Verdicts in plain words

    Discarded, superseded, reference, my pick. No scores, no percentages of quality. If something is just an impression, I say it is an impression.

  5. 05

    What I learned, with receipts

    Three to five learnings per chapter, and each one links to the attempts that back it. If I can't point at the evidence, it isn't a learning yet.

What I use

Nothing exotic. The interesting part was never the tool.

  • ComfyUI, locally

    My own machine. Every image here came out of a graph I can open again.

  • Krea 2 and its variants

    Turbo, Niji for anime, Comics. Plus LoRAs with their weights written down, because the weight is usually the whole story.

  • Seeds and steps, on the record

    Each attempt shows its model, its seed and its steps. Two images with the same seed and a different flaw say more than either one alone.

  • Upscalers, compared

    Lanczos, AnimeSharp and SeedVR2. I put them on the same image and looked up close instead of trusting the name.

House rules

  • Failures are published. A logbook that only shows wins is a portfolio.
  • Collaborations are credited by name, always.
  • Prompts stay in English, exactly as they were used.
  • The clips and the images I picked carry a small watermark; the contact sheets don't.

Where it all goes

See the experiments