SDXL native resolution picker
Stable Diffusion XL produces its cleanest results at the resolutions it was actually trained on. This picker lists every official SDXL training bucket, each landing near the model’s one-megapixel target, with exact width, height, simplified aspect ratio, and megapixel count — and one click copies the dimensions ready to paste into your SD interface.
Why native resolutions matter
SDXL was trained using a technique called aspect-ratio bucketing. Rather than resizing all training images to a single fixed square, the data was grouped into a set of width/height pairs that all sit close to roughly 1,024 × 1,024 pixels (about one megapixel). Each bucket represents a specific canvas shape the model has seen thousands of times. Generating at one of those buckets means the model frames subjects the way it learned to frame them.
When you generate at a size that does not match a trained bucket, the model has to extrapolate to a canvas shape it has never seen. The quality cost depends on how far you stray:
- Too small (below ~768 on the short edge): soft, low-detail output — SDXL is not SD 1.5 and does not work well at small sizes.
- Too large in one pass (above ~1.5 megapixels): duplicated subjects, twin heads, repeated patterns — the model tile-fills rather than framing a single coherent scene.
- Off-ratio (not a trained bucket): stretched proportions, awkward framing, subjects misplaced toward edges.
The SDXL training buckets
All official SDXL training resolutions sit near one megapixel. A selection of the most commonly used:
| Aspect ratio | Width | Height | Typical use |
|---|---|---|---|
| 1:1 (square) | 1024 | 1024 | Portraits, icons, general use |
| 4:3 | 1152 | 896 | Landscape, wide subjects |
| 3:4 | 896 | 1152 | Portrait orientation, tall subjects |
| 16:9 | 1344 | 768 | Widescreen, cinematic framing |
| 9:16 | 768 | 1344 | Mobile / story format |
| 3:2 | 1216 | 832 | Photography landscape ratio |
| 2:3 | 832 | 1216 | Photography portrait ratio |
The picker lists all of these plus less common intermediate buckets for fine-tuned aspect ratios.
The right workflow: generate native, then upscale
The temptation to generate directly at 2048 × 2048 to get a larger image is almost always counterproductive. Stick to the native resolution workflow:
- Generate at a native SDXL bucket — get the composition, lighting, and subject right at ~1 megapixel.
- Upscale with hires fix or a dedicated upscaler — tools like ESRGAN, RealESRGAN, or the built-in hires fix in AUTOMATIC1111 add genuine detail in the second pass rather than simply enlarging existing pixels.
The second pass at 2x upscale with 0.35–0.5 denoising strength resolves facial features, hair strands, and background detail that SDXL cannot place correctly in a single direct-to-large-canvas generation.
Tips for choosing a resolution
- Portrait subjects: 896×1152 or 832×1216 are the most natural for people and vertical compositions.
- Landscape shots: 1216×832 or 1344×768 frame scenery naturally in horizontal format.
- Cinematic 21:9 widescreen: is not a native bucket — approximate it with 1344×768 and crop after upscaling.
- Square (1024×1024): the safest default and the most thoroughly trained single bucket; when in doubt, start here.