Generating AI art at the wrong dimensions causes stretched faces, awkward crops, duplicated limbs and wasted compute. Picking model-friendly sizes fixes most of it. This guide covers the native resolutions and best aspect ratios for the major tools — and our free AI image size calculator gives you ready-to-paste dimensions for any ratio.
Diffusion models are trained at specific resolutions. Generate far from those and quality drops or you get artifacts (the classic "two heads" problem comes from going too wide/tall). Each model also works in latent blocks, so dimensions should be multiples of 8 (ideally 64).
| Model | Native (1:1) | Note |
|---|---|---|
| Stable Diffusion 1.5 | 512 × 512 | Upscale afterward for detail |
| SDXL | 1024 × 1024 | Best results near ~1MP total |
| Midjourney v6/v7 | ~1024 base | Set shape with --ar, then Upscale |
| Ratio | Use | Dimensions |
|---|---|---|
| 1:1 | Avatars, posts | 1024 × 1024 |
| 16:9 | Wallpaper, video | 1344 × 768 |
| 9:16 | Phone, Shorts/Reels | 768 × 1344 |
| 3:2 | Photography | 1216 × 832 |
| 4:3 | Classic | 1152 × 896 |
Why do multiples of 64 matter? Diffusion models operate on a downsampled latent grid; sizes aligned to 8/64 map cleanly and avoid edge artifacts.
What's the best size for SDXL? Aim for about one megapixel total — 1024×1024 for square, or the ratio equivalents above.
Can I just generate at 4K? Usually not in one pass — generate at the native budget and upscale for high resolution.