DotStories BlogOpen DotStories →

A 12B image model now runs fully in your browser

Running a big image model usually means installing something first: ComfyUI, Python, the right drivers, a launcher. That's an afternoon gone before you see a single image.

Now you can skip all of it. Krea 2 Turbo, a 12-billion-parameter image model, runs fully in your browser.

No install, no server, no re-download

Try it

  1. Open the Krea 2 page in Chrome or Edge.
  2. Pick your ComfyUI models folder.
  3. Write a prompt and press Generate.

You need these three files in that folder, the same ones ComfyUI uses:

Have the Dark Fantasy Kit's loras/krea2_darkbrush.safetensors too? Pick it in the LoRA menu and get the same brush style as the Dark Fantasy Thumbnail dot.

How fast is it?

On an RTX 4080, one image takes about 41 seconds at 512×512 and about 136 seconds at 1024×1024, including loading the model onto the GPU. That's slower than ComfyUI for now. We're working on it.

You'll need a desktop GPU with about 16 GB of memory. Not sure yours can do it? Run the probe first: it checks your browser and GPU in a few seconds.

Same result as ComfyUI

Browser AI usually stops at 4 GB, and Krea 2 needs about 18. So we wrote our own small engine: it moves the model from your disk to your GPU one layer at a time, and unpacks the fp8 weights right on the GPU.

Every step is checked against ComfyUI's own code, so you get the image ComfyUI would make, not a rough copy.

What's next

Krea 2 has its own license. Read it and follow it before you use the model.

Open source

The engine is MIT-licensed and open to everyone: github.com/dotstories/web-engine. Tried it on your GPU? Share your time on Discord.

Make an image with Krea 2 in your browser →

← All posts