Image-Gen Resolution Picker (SDXL · Flux · Midjourney) | RECATOOLS

How to Use the Image-Gen Resolution Picker

Pick your model

Choose SDXL, SD 1.5, Flux or Midjourney from the selector. Each model was trained at different sizes, so the "right" resolution is model-specific — there is no single correct number.

Find your aspect ratio

The grid lists every recommended resolution with its aspect-ratio label — square, portrait, landscape, wide and tall. Scan for the shape you want for your image.

Click to copy the dimensions

Tap any card and the exact WIDTHxHEIGHT is copied to your clipboard, ready to paste into Automatic1111, ComfyUI, Forge, or a Midjourney --ar note.

Generate, then upscale

Render at the native size first for clean composition, then upscale to your final delivery resolution. Generating off-bucket to "save an upscale" is what causes most artifacts.

Why Resolution Is the First Decision in Image Generation

Diffusion models have a "native" resolution — and it is not a suggestion

Every diffusion model is trained on images at specific sizes, and it learns the structure of the world at that scale. SD 1.5 was trained almost entirely on 512×512 crops; SDXL was trained on a set of carefully chosen "buckets" centred on roughly one megapixel, like 1024×1024, 1152×896 and 1216×832; Flux is comfortable across a wider range up to about two megapixels but still prefers dimensions that are multiples of 16; and Midjourney renders a base image near one megapixel and then upscales. When you ask a model to generate at a size it never saw in training, it has to improvise the large-scale composition — and that is exactly where things break.

The classic failure is the "double-head" or "twinning" artifact: ask SD 1.5 for a 512×1024 portrait and you will often get two torsos stacked on top of each other, because the model only knows how to lay out a single figure inside a roughly square frame and is forced to tile its knowledge to fill the extra height. SDXL is far more robust thanks to its bucketed training and conditioning on the original image size, but it too degrades when you wander off its known aspect ratios — fine detail smears, faces lose coherence, and horizons duplicate. Picking a documented resolution is the single cheapest way to avoid all of this before you ever touch a prompt.

"Off-resolution generation is the most common, least-diagnosed cause of bad AI images. Fix the size first, then blame the prompt."

Buckets, megapixels and the multiple-of-64 rule

There is also a hard technical reason the numbers look odd. A diffusion model works in a compressed latent space, and the VAE that compresses the image downsamples by a factor of eight. That means both your width and height should be divisible by 8 — and in practice tools nudge you to multiples of 64 — or the latent grid does not line up cleanly and you get edge artifacts. This is why the SDXL buckets are values like 1344×768 rather than a tidy 1350×760: each one is divisible by 64, keeps the total area near one megapixel, and gives a useful aspect ratio. Flux relaxes this slightly to multiples of 16 and allows a larger area, which is why it can comfortably do 1440×1024 where SD 1.5 cannot.

The practical workflow that falls out of this is simple. Choose the model you are actually running, pick the bucket whose aspect ratio is closest to what you want, generate there, and only then upscale — with an ESRGAN-style upscaler, a tiled diffusion pass, or your platform's built-in upscale — to reach a billboard-sized final file. Generating directly at 4K to skip the upscale almost never works on SD-family models: you exceed the resolution the model understands, lose global coherence, and spend far more compute for a worse result. Midjourney is the exception in that it abstracts all of this away behind --ar, but even there extreme ratios crop the composition the model planned, so the bucket-aware mindset still pays off. This picker exists so you never have to memorise these numbers: choose a model, read the aspect-ratio labels, and copy a size that the model was actually built to produce.

10 Facts About Image-Gen Resolutions

01

SD 1.5 was trained at 512×512 — push much past that on one side and you get duplicated heads or bodies.

02

SDXL trains on ~1-megapixel "buckets" like 1024×1024, 1216×832 and 1344×768, not one fixed size.

03

Width and height should be divisible by 8 (often nudged to 64) because the VAE downsamples by eight.

04

The "twinning" artifact — two figures where you wanted one — is usually an off-resolution symptom, not a prompt problem.

05

Flux is happy up to ~2 megapixels but prefers dimensions that are multiples of 16.

06

Midjourney has no fixed buckets — you set any ratio with --ar and it renders near 1 MP, then upscales.

07

SDXL conditions on the original image size, which is why it tolerates more aspect ratios than SD 1.5.

08

Generate at native size, then upscale — going straight to 4K on SD-family models loses global coherence.

09

The odd numbers (1344, 832, 1216) exist so each side is divisible by 64 while keeping the area near 1 MP.

10

Picking a documented resolution is the cheapest fix in image gen — size first, prompt second.

Frequently Asked Questions

Use one of SDXL's trained buckets: 1024×1024 for square, 832×1216 or 896×1152 for portrait, 1216×832 or 1152×896 for landscape, and 1344×768 or 1536×640 for wider shots. They all sit near one megapixel and are divisible by 64. Pick the one closest to the aspect ratio you want and copy it from the grid above.
That "twinning" artifact almost always means you generated at a resolution far from the model's native size — most often a tall portrait on SD 1.5. The model only knows how to compose a single figure inside its trained frame, so it tiles its knowledge to fill the extra space. Generate at a documented size and then upscale, and the problem usually disappears.
Diffusion models work in a compressed latent space, and the VAE that compresses the image downsamples by a factor of eight. If width or height is not divisible by 8 the latent grid does not line up cleanly, producing edge artifacts. Most tools round to multiples of 64 for safety, which is why the recommended numbers look like 1344 and 832 rather than round figures.
On SD 1.5 and SDXL, no — you exceed the resolution the model understands and lose global composition while burning far more compute. The reliable path is to generate at the native bucket size for clean structure, then upscale to your final delivery resolution. Flux tolerates larger sizes, and Midjourney handles upscaling for you.
SD 1.5 is native to 512×512. For portrait or landscape, 512×768 and 768×512 are the safe choices; you can nudge to 640×768 or 768×640, but pushing one side much past 768 invites duplicated subjects. Generate small and upscale rather than asking for a large frame directly.
Flux is more flexible: it handles a wider range up to roughly two megapixels and is happy with dimensions like 1440×1024 that SDXL would struggle with. The main rule is to keep both sides divisible by 16. Around one megapixel at a clean ratio remains the sweet spot for speed and coherence.
Midjourney abstracts pixel sizes away: you set the shape with the --ar flag (for example --ar 16:9), it renders a base image near one megapixel, and then upscales. The pixel figures in the grid above are the approximate dimensions each ratio produces — handy when you need a concrete number, but you control the output with the ratio, not a width and height.
It copies the plain dimensions in WIDTHxHEIGHT form, for example 1216x832. Paste it into the width and height fields of Automatic1111, ComfyUI, Forge or your platform of choice. Nothing is sent anywhere — the copy happens entirely in your browser.
No — they are the documented, well-behaved ones. You can always try other sizes, especially with SDXL and Flux which tolerate variation. The grid simply gives you the resolutions that the models were built around, so you start from a size that is known to produce coherent images.
Completely free, with no account and no limits. The resolution table is built into the page, so it loads instantly, calls no AI and no server, and stores nothing — selecting a model and copying a size all happen locally in your browser.

Related News

You may be interested in these recent stories from our newsroom.

No related news yet for this tool. Our editorial team publishes new pieces every week.

Browse all news →

Image-Gen Resolution Picker