A/B Challenge Preference Test Krea2-Turbo vs. Qwen 2.1 Base
retained text
I thought it might be fun to do a little A/B challenge/preference test. I generated 5 prompts on ChatGPT that should work to the strengths of each model around the same concept and then used the default Comfy templates with a couple tweaks (no LoRA, just disabling some nodes for prompt enhancers or k/v caching, fixing bugs like not randomized seed). I used the default fp8 Krea2-Turbo model and encoder and the BF16 Qwen 2.1 model and encoder. Krea2 images used the default 8 step settings since being distilled doesn't give a tone of wiggle room with just the turbo and generally seem to be the best settings for the majority of images. Qwen 2.1 not being distilled uses a mix of Heun and DPMPP\_S2\_Ancestral at 40 steps at a CFG of 4.0, a couple images used linear\_quadratic for the scheduler, but the rest used simple. Some Qwen 2.1 images had negative prompts as well. The images are random for which model is A/B. Post guesses/preferences and I'll post the answer key later and respective prompts.
document fetched 2026-10-09 00:33 · canonical https://reddit.com/r/StableDiffusion/comments/1x11q9m/ab_challenge_preference_test_krea2turbo_vs_qwen