T2I Model Comparison

What it is

An open comparison of text-to-image models over a single prompt set. Not a selection of good-looking examples and not a judgement by eye, but a full run: every prompt through every model, with all results published.

600 prompts, 17 models, 10,200 images.

Why it exists

Comparisons of generative models are usually built on a handful of striking images picked by the author. They tell you nothing about how a model behaves on a boring prompt, on a long description, or on a specific kind of scene — which is exactly what decides whether it is usable for work.

The only way to answer that is to run the same prompt set through every model and publish the result in full, failures included.

What I did

Built the prompt set, ran it through 17 models on my own GPU stack, published the dataset and an interactive comparison. The pipeline, the run and the publication are entirely my work.

The link to KRT Studio

The whole set was generated through KRT Studio — as batch generation inside the studio, not as a separate script.

That is the real stress test of the platform. The first run produced 6,600 images across eleven models. The set then grew to seventeen, bringing the total to 10,200 images and around 72 hours of generation. Jobs ran in the background, the window could be closed, and state survived a restart.

Generating one image is easy. Carrying a run of dozens of hours without losses, switching models and staying within available VRAM, is a different class of problem — and it was solved at the architecture level: batch locking, a status registry, state reconciliation.

What came out of it

The dataset — all 10,200 images with their prompts and generation parameters, available for your own analysis.

The interactive comparison — pick a prompt and put the models’ results side by side. That matters more than any single number: instead of the claim “model A beats model B”, a person looks at a specific prompt and decides for themselves.