Xingyu-Zheng/MrFlow — reverse-engineered prompt

Build me a small Python project that recreates the MrFlow idea for faster text to image generation. I want it to take a prompt, generate a low resolution image first, upscale it with Real ESRGAN x2, re encode it, add the right amount of noise, then do a very short high resolution refinement pass so it can get close to normal sampling quality but much faster.

Please make the main scripts work for Qwen Image and FLUX.1 dev like in the repo, with easy checkpoint path placeholders, a simple 12plus1 and 20plus1 mode, and output folders that save the low res image, the upscaled image, and the final refined image. Keep the code minimal, readable, and runnable, with shared helper code and a clear readme that explains what to edit and how to run it.

If extra examples for FLUX.2 Klein, Z Image, or Pi Flow are easy, include them as optional examples only. Look up the current docs and paper if needed, and use the Real ESRGAN source and weights the readme points to.