>>2978215
When inpainting, you can choose the target resolution. Your inpaint area gets upscaled to that size, which lets the model clean up details it otherwise couldn't. For images where the detail (face) is this far off from the camera, you either need higher resolution or you need to keep the inpaint area (plus padding) very small, or a mix of both. Also helps if you cut down the prompt to just 1girl and features of her face+hair. Could also be that a lora you use is trained on messy eyes, in that case you might try lowering its strength or disabling it while inpainting.

I didn't get the second part, how do backgrounds get increasingly worse when you don't inpaint?