A dim tunnel full of graffiti is a great place to find inspiration for your next perfect portrait.
All images are generated with Colab notebook included in the Quick Start Guide.
Positive prompt:
100mm, cover photo portrait, [liza soberano: emma watson:0.7], skirt, stocking,4 k, (dimly lit urban tunnel filled with beautiful graffiti:1.1), liam wong, directional lighting
Negative prompt:
| CFG scale | 7 |
| Image size | 512×704 |
| Sampling steps | 30 |
| Seed | -1 |
| Face restoration | None |
| Sampling method | Euler a |
| Model | F222 |
| Recommended GUI | AUTOMATIC1111 |
Technical notes
- I intended to get a full body realistic portrait. The keywords photo portrait and the portrait image size (512×704) helped. The keyword skirt and stocking helped to make the legs visible.
- Perhaps due to the keyword tunnel, the initial prompt generated a subject too far away. Close up keyword made it too close, making the tunnel not visible. Finally I found the right balance with the keyword cover.
- It was hopeless to generate a good face with that small face size. All faces were fixed with a second inpainting step.
- To finely control the look of my model, I used keyword blending to mix two faces. I chose Liza Soberano for an asian look, and blend with Emma Watson to make the facial feature sharper.
- Keyword stocking is used to make the lower body more interesting.
Inpainting settings
You will get pretty bad faces because Stable Diffusion is not good at generating small faces.
Inpaint with original prompt and model. The important parameters are
- Mask: The face only
- Prompt: (same as original)
- Mask content: Original
- Inpaint at full resolution
but with full resolution on and face restoration.
Below is a screenshot of the full setting.