generative models are funny. every time a new advanced model comes out like qwen, it makes previous models look incompetent. but then we immediately start testing their limits and run into a completely new set of problems.

qwen is qualitatively superior to previous open models at prompt following and text. but it still can have lots of errors when you start pushing it. it doesn't help that I have to run it at fp8 or quantized, degrading quality a bit. and each gen takes ~2 minutes on my machine, so iterating on a prompt to get better results takes much longer too (unless iterating at lower steps with a distill).

>>106253751
a couple seconds? more like 12 minutes due to how slow this thing is.

>>106253822
>mgsv was ten years ago
bros...