Kandinsky 2.1: open-source txt2img generator, image blending, beats SD on FID

https://github.com/ai-forever/Kandinsky-2

by famouswaffles • 3 years ago

6 2 3 years ago

After testing it out, this is a much better base model than any of the Stable Diffusion models out (incl XL Beta).

Understands prompts better and generates much better looking images for the same prompt. And the image mixing/blending is far better than any other open source alternative ( lambda mixer, versatile diffusion etc)

You can try it out on

https://huggingface.co/spaces/ai-forever/Kandinsky2.1

https://fusionbrain.ai/diffusion

brutusurp• 3 years ago

Just tested it out. The Kandinsky 2.1 model does render high-quality images accurate to input prompt. Each iteration of a prompt produces more robust interpretation; exciting. Still working through the relationship between the positive and negative prompts.

There is a bit of a gap in understanding metaphor and reverse polarity. For instance if I include in the prompt "move away from negativity" Kandinsky produces images of sad faces (e.g., mouth curved down, downward-focused eyes). However I expected there to be an inferred "towards positivity" interpretation (e.g., eyes closed/looking forward or upward, and mouth relaxed/smiling). Trying to see how much this outcome can be achieved leveraging the negative prompt field.

Overall I'm happy with the results and will continue to use it. Thanks for sharing!

Kandinsky 2.1: open-source txt2img generator, image blending, beats SD on FID

Related Stories