Realism Qwen Image 2.1 Edit LoRA — v2
Turns a polished AI image (smooth skin, flat light, blurred background) into something that looks like a real phone photo. It is an edit LoRA: give it your AI image as the input image plus the prompt below.
How to prompt
Use this exact prompt — the LoRA was trained on it and nothing else:
convert this into a real phone photo
Don't describe the image or add style words; the input image carries the content.
Recommended settings
| setting | value |
|---|---|
| LoRA strength | 0.5–0.75 (start at 0.6). 1.0 works but can make skin look prickly and change background details such as windows. |
| steps | 25 |
| CFG | 1 |
| resolution | around 1 megapixel, keep the input's aspect ratio |
Known faults
- At high strength: gritty/prickly skin, eyes can look less natural, scene objects (windows, frames) can shift or duplicate.
- Lower the strength before anything else.
Training (for reference)
Fizgig, Qwen Image 2.1 Edit preset (rank 8, adaptive LR), 30 pairs, 330 steps. Pairs were made in reverse: real photos were pushed into an exaggerated "AI look" and the LoRA learns to undo it.