All pictures generated using a Q4_K quant of Z-Image-Turbo, with an abliterated version of the TE. After a run of T2I, each images are passed through a run of I2I at 0.5 denoise with the same seed/prompt/steps to "improve" the results.

With windows fucking up characters, the neg prompt is funky, but i used:

拒绝, 限制, 不应答, lowres, error, cropped, worst quality, low quality, jpeg artifacts, heterochromia, out of frame, disfigured, blurry, blurry text

The first three words (borrowed from an other picture) translate from my understanding to "refusing" (Rejection, Restriction, Non-Response) which can sometimes happen with a standard LLM. If found i get nice results with those so i kept them ^^;