Misleading Claims, Outdated Outputs: Kandinsky 3.1 Feedback, Worst model I Ever Used!

I’ve never encountered such a misleading, lying model before. I installed Kandinsky 3.1 and spent the whole evening generating images. Kandinsky 3.1 is by far the worst of all the models I’ve tried. Not only is it unbearably heavy—requiring 50 steps per generation (though my GPU handles that easily)—but it also produces outrageously poor-quality results. It has terrible prompt fidelity, and the images look like they were generated by the first diffusion models from five years ago. All the “promo” claims by the developers that Kandinsky beats every other model of its time are lies. All my prompts—both English and Russian—ended up as trash. Even their own demo prompts (“A bear in a Russian national hat with a balalaika,” “Maximally realistic portrait of a jolly old gray-haired Rastaman with wrinkles around his eyes and a crooked nose in motley clothes,” etc.) give far worse results than any model before SD 1.5. By default, it only generates 1024×1024 and the quality is so bad already. I’m using Kandinsky3.1 (45 GB) and flan-ul2-text-encoder (16 GB), and the results are exactly as described above.