TEST RECORD · ← all field reports

FIELD REPORTS/IMAGE/ISSUE #108

Make Krea 2 images twice as fast without losing fine detail

We covered Krea 2 Turbo in issue 63 as the fast way to sketch image ideas. One person with a single RTX 3090 spent three weeks making it faster still, and published every measurement along the way.

MODELKrea2-Turbo-Distill-4step-LoRA
PUBLISHEDSeptember 14, 2026
READ TIME3 min
TESTED BYNeural Expedition
CATEGORYIMAGE

Field notes

01What it does

Krea 2 Turbo needs about eight passes to produce a clean image. Cut that to four and it gets blurry: fine texture goes soft and edges lose definition. This adapter, a LoRA that loads on top of the unchanged base model, was trained to put that detail back at four passes. The author reports an end-to-end render at 1024 by 1024 dropping from about 89 seconds to about 55 on the test machine, with fine-texture energy at 1.03 times the eight-step original and saturation and skin texture within a few percent of it. A vision-language judge shown the eight-step and four-step renders side by side preferred the original on only 6 of 45 pairs.

There is one dial. Strength 1.0 is the trained setting; if a stylised or high-contrast prompt comes back looking too crunchy, 0.75 softens the correction without changing the picture itself. It was trained at twelve resolutions and aspect ratios from 512 square up to 1440 square, so portrait and landscape framings are covered.

The repo is unusually honest about how it was made: which training stretches made things worse, why the shipped weights are a running average rather than the last checkpoint, and every comparison sheet at a fixed seed. Nothing here is Krea's official work; it is one person's project, and the numbers are the author's own.

02How to try it

A community demo Space runs the adapter on Krea 2 Turbo, so start there: type a prompt with visible texture in it, a freckled portrait, tweed, wet stone, and compare the four-step result with what you remember from the Krea 2 demo. To use it in your own setup, the repo ships two files with identical weights: one for diffusers and one for ComfyUI, plus a ready ComfyUI workflow that uses stock nodes only. Load the Krea 2 Turbo base as usual, add the LoRA, set four steps and guidance to zero, and everything else stays the same. Each step you drop is a step off every image in a batch, so this pays off most when you generate a lot.

03Caveat

The adapter inherits the Krea 2 community license, including its revenue threshold for commercial use and the content-filtering duty for public deployments. Speed gains are relative to the eight-step baseline on the author's hardware; the base model itself is still large, so a small GPU stays slow. This is an accelerator for a model you already run, not a new model.

04What you can do with it

  • Turn a two-minute batch of concept images into a one-minute batch with the same prompts and seeds.
  • Iterate on product and campaign mockups with a faster preview loop before the final render.
  • Get usable four-step images on an Apple Silicon Mac, where every skipped pass matters more.
  • Run side-by-side tests with the strength dial to see exactly what the adapter adds.

Try the demo

View model page

Read this issue on neuralexpedition.com →