wlejon/krea2-research

This repository is a research artifact rather than a finished product. It documents an investigation into Krea 2, an AI image generation model, looking for internal controls that can be adjusted in real time to change the resulting image within a couple of seconds, instead of having to write a new text prompt and regenerate from scratch. The full write up of what was tried and what was found lives in two documents in the repository, one covering the plan and one covering the findings. To explore this, the project includes a small local web interface built with Gradio, a Python tool for building quick interactive demos. It gives you sliders for things like composition, color, lighting, rendering style, and mood, plus a dial to shift an image between a literal, photographic look and a more stylized, painterly one. There are also slots for creating your own custom sliders from a pair of text descriptions or a pair of images. Every slider updates the generated image when you let go of it, usually within a few seconds at a small image size. A separate tab lets you apply one of these adjustments to just part of an image rather than the whole thing, using a mask. Each slider also has a small button that automatically sweeps it across its entire range, regenerating the image at each step and stitching the results into an animated gif so you can see the full effect of that control at a glance. The repository includes a few example videos and gif sweeps generated this way. Overall, this is a lab notebook for people studying how these image models can be steered, rather than a tool meant for general end users.