The Crafting Table for Your Brain

Krea dropped a prototype today where you build a custom LoRA by dragging images into a grid and merging them on top of each other.

That's it. That's the whole thing. Drop images, smash five together, get a model trained on your concept — character, object, aesthetic, whatever you're after. No Python, no config files, no cloud credits you have to squint at before you know if you can afford to finish the thought.

The merge threshold is five. Not four, not ten — five, which is arbitrary and also completely correct, the way all game crafting recipes feel arbitrary until you've internalized them and then they feel like physics. The UI is teaching you that five images is a concept. Fewer than that is a vibe.

Text support is coming. Right now it's images only, which is the right call — text-to-concept is a harder UX problem and they didn't need to solve it to ship the interesting part.

What they're building toward, and what "full diffusion UI" means when they say they're getting close, is a canvas where the entire pipeline collapses into direct manipulation. You don't configure a model. You don't pick a checkpoint. You drag your references into a pile, the pile becomes a model, and then you use the model. The abstraction is the action.

This is the thing that kills the command line for image generation — not better outputs, not cheaper compute, but the moment where training a LoRA feels indistinguishable from arranging files on a desktop. The cognitive overhead drops to zero and suddenly the tool is for everyone who has opinions about what things should look like, which is everyone.