A/B Blind Taste Test - Krea2-Turbo vs. Qwen-Image 2.1 Turbo
Since Qwen 2.1 Turbo dropped suddenly right after my last taste test, figured might as well do one for it as well.
Instructions:
Each image is either Krea2-Turbo or Qwen 2.1 Turbo, using the exact same prompt.
You do not know which one is which, it could be either A or B on each individual image.
Pick your favorite based on preference and list them.
You can also guess which one of each is Krea2 or Qwen 2.1 Turbo.
The answer key will be posted 24 hours after posting.
Technical Setup
This time all images were generated at 2MP, using 8 steps. Krea2-Turbo was the fp8 model and text encoder, but Qwen was BF16 for both model and text encoder.
Krea2 used Euler-Simple, where as the still less understood Qwen Turbo I went with Linear Quadratic (I know there is an official sigma curve for it, but its close enough) and since Euler seems also not be a good sampler, I went with the tried and trusted Heun, which gave comparable inference times to Krea2 (Qwen 2.1 Turbo is about twice as fast as Krea2-Turbo on Euler).
I did test (and have in the past) different samplers and schedulers in Krea2 but overall they generally seem to all be equal, so I did not differ from the standard Euler-Simple.
Prompts:
All prompts are the same for both models. They were generated in ChatGPT, expanding off basic ideas. Some of the expanded prompts were not great and were tweaked manually when both models demonstrated consistent inconsistencies with the prompt idea, and reran on both models. Each prompt was ran up to 10 times to find a suitable generation.
The prompts for each image are here: https://paste.sh/yDAHasri#TveZa0FgYZ-UJ89Exlvpbelm