Stable Diffusion SDXL, the popular AI image generator, has a tendency to create duplicate elements, entities, and persons in images when not used with the recommended image sizes.
This occurs because Stable Diffusion is trained on square image sizes and fine-tuned on official sizes.
Best results usually happen with 1:1 or 1:2 size but with SDXL now we have more options.
Official Recommended Sizes
The Stable Diffusion team recommends using these aspect ratios for best results:
- 1024 x 1024 (1:1 Square)
- 1152 x 896 (9:7)
- 896 x 1152 (7:9)
- 1216 x 832 (19:13)
- 832 x 1216 (13:19)
- 1344 x 768 (7:4 Horizontal)
- 768 x 1344 (4:7 Vertical)
- 1536 x 640 (12:5 Horizontal)
- 640 x 1536 (5:12 Vertical, the closest to the iPhone resolution)
When generating images, it’s best to stick to these recommended sizes.
These sizes are the fine-tuned sizes for SDXL. after the model training over model fine tuned for this sizes so it will give better results. some of them works stable some dont.
Not all sizes are perfect. Be sure to stick to pixel by pixel same recommended sizes. If you need bigger size just use the upscale don’t generateupscale directly that will be a problem.
Going too far outside the recommended sizes can cause Stable Diffusion to ‘invent’ details that aren’t part of the original image.
My Favorite Generation Size
1344 x 768 is my most generated size. It works best for me and it works really well with almost all types of prompts. Trust me I tested lots of prompts 😉
Why this size better than others?
Well because it is almost an HD ratio and when you upscale/multiply with 1.43 it gives you 1920×1096 PERFECT.
Perfect for HD image generation.
If you like generating HD images like me you don’t have another size for SDXL this is the only way 🙂