Best Image Size for Stable Diffusion

The best image size for Stable Diffusion is the model’s native training resolution: 512 x 512 for SD 1.5, 1024 x 1024 (about 1 megapixel) for SDXL and SD 3.5, and 1 to 4 megapixels for the FLUX family. Generate at that size, upscale afterwards, then crop to the dimensions your destination needs. For a WordPress featured image that means a 1024 x 1024 or 1344 x 768 render, upscaled 2x and cropped to 1200 x 628.

Asking the model for an odd size directly, such as 1200 x 628 on SD 1.5, is the single most common cause of doubled heads, repeated limbs and smeared backgrounds. The model has never seen that canvas, so it tiles what it knows. The fix is cheap: render natively, then resize.

This guide lists the native sizes for every current Stable Diffusion and FLUX model, the aspect-ratio presets that work, the exact pixel sizes WordPress, WooCommerce and BuddyPress expect, and the compression step that stops a 6 MB PNG from wrecking your Core Web Vitals.

How Stable Diffusion Image Sizes Work

Diffusion models learn on images at a fixed pixel budget. SD 1.5 was trained at 512 x 512. SDXL and SD 3.5 were trained around 1,024 x 1,024, with a set of “bucketed” aspect ratios that all add up to roughly one megapixel. FLUX models are trained across a range from about 0.1 to 2 megapixels, and FLUX.2 extends that to 4 megapixels.

Generating inside that budget gives you the composition quality you see in demos. Going well above it makes the model repeat subjects because it treats the extra space as a second canvas. Going well below it loses detail and produces mushy faces. The total pixel count matters more than the exact width and height, which is why 1344 x 768 works fine on SDXL even though it is not square.

There is also a hard rule: width and height must be divisible by 8 (SD 1.5, SDXL) or 16 (SD 3.5, FLUX) because the latent space is downsampled by that factor. Most interfaces enforce it, but if you type dimensions by hand, round to the nearest multiple of 64 to be safe.

Native Output Sizes by Model

Model	Native square	Landscape presets	Portrait presets	Notes
SD 1.5	512 x 512	768 x 512, 640 x 448	512 x 768, 448 x 640	Oldest, fastest, huge LoRA library; needs upscaling for anything web-sized
SDXL 1.0	1024 x 1024	1152 x 896, 1216 x 832, 1344 x 768, 1536 x 640	896 x 1152, 832 x 1216, 768 x 1344, 640 x 1536	Still the most common base for fine-tunes and LoRAs
SD 3.5 Large / Medium	1024 x 1024	Same 1 MP buckets as SDXL	Same 1 MP buckets as SDXL	Medium handles 0.25 to 2 MP; free under the Community License below $1M revenue
FLUX.1 dev / schnell	1024 x 1024	Anything from 0.1 to 2 MP, e.g. 1536 x 1024	e.g. 1024 x 1536	Best text rendering in the open-weight group
FLUX.2 dev / klein	1024 x 1024	Up to 4 MP, any aspect ratio, e.g. 2048 x 2048	e.g. 1536 x 2048	32B parameters; needs a 24 GB card or 4-bit quantisation

Always generate at the model’s native resolution and then upscale afterwards. Real-ESRGAN, 4x-UltraSharp and the SUPIR and Ultimate SD Upscale nodes in ComfyUI all produce far better results than asking the base model for a larger canvas. If you use Automatic1111 or Forge, the “Hires. fix” option does the same two-stage job in one click.

Best Image Sizes for WordPress

WordPress uses different image sizes for different purposes. Here are the dimensions to crop to once you have a clean upscaled render:

Usage	Recommended size	Generate at (SDXL / SD 3.5 / FLUX)	Notes
Featured image	1200 x 628 px	1344 x 768, upscale 2x, crop	Works for blog cards and social sharing (OG image)
Blog header / hero	1920 x 600 px	1536 x 640, upscale 2x, crop	Full-width banner images; leave empty space for headline text
Thumbnail	300 x 300 px	1024 x 1024, downscale	WordPress default thumbnail
Medium size	768 x auto	1024 x 1024 or 1152 x 896	Used in content body
WooCommerce product	800 x 800 px (1200 x 1200 for zoom)	1024 x 1024, upscale 2x	Square format, supports zoom
BuddyPress cover image	1950 x 450 px	1536 x 640, upscale 2x, crop	Profile and group cover photos
BuddyPress avatar	300 x 300 px	1024 x 1024, downscale	Square, auto-cropped
Open Graph (social)	1200 x 630 px	1344 x 768, upscale 2x, crop	Facebook, LinkedIn, X cards
Pinterest pin	1000 x 1500 px	832 x 1216, upscale 2x, crop	2:3 portrait

Two WordPress defaults to keep in mind. Uploads wider or taller than 2560 px are scaled down automatically (the “big image” threshold), so there is no benefit in uploading a 4096 px render. And the media library’s default sizes (150, 300, 768, 1024 px) are generated from whatever you upload, so upload the largest size you actually display and let srcset do the rest.

Generating the Right Size in Stable Diffusion

The workflow we use for every AI image on this site:

1. Pick the closest native bucket. For a 1200 x 628 featured image (1.91:1), the nearest SDXL bucket is 1344 x 768 (1.75:1). For a 1920 x 600 hero (3.2:1), use 1536 x 640 (2.4:1) and accept a wider crop.
2. Generate at that size. Use 1344 x 768 for SDXL, SD 3.5 or FLUX. Batch four seeds and keep the best one.
3. Upscale 2x. Real-ESRGAN or Ultimate SD Upscale takes 1344 x 768 to 2688 x 1536. Use a low denoise (0.2 to 0.35) so the upscaler adds texture without changing the composition.
4. Crop to target dimensions. Canva, Photoshop, GIMP or a one-line ImageMagick command: magick in.png -resize 1200x628^ -gravity center -extent 1200x628 out.png.
5. Compress and convert. Export as WebP at quality 80 (or AVIF at 60) before uploading. A typical 1200 x 628 featured image should land under 120 KB.

Aspect Ratios That Work

• 1:1 (1024 x 1024): product images, avatars, thumbnails, Instagram
• 16:9 (1344 x 768 or 1536 x 864 on FLUX): blog headers, video thumbnails, hero banners
• 1.91:1 (crop from 1344 x 768): Open Graph images, featured images
• 4:3 (1152 x 896): general content images
• 3:4 / 2:3 (896 x 1152 / 832 x 1216): Pinterest pins, portrait shots, mobile-first heroes
• 21:9 (1536 x 640): ultra-wide banners and BuddyPress covers

Worked Example: a Featured Image in ComfyUI

Load the SDXL or FLUX.1 dev checkpoint. Set Empty Latent Image to 1344 x 768, batch size 4. Write the prompt with the subject on one side (“subject on the left third, plain background on the right”) so you have room for a title overlay later. Sample at 25 to 30 steps. Pass the winner through an Upscale Image (using Model) node with 4x-UltraSharp, then a Scale Image node at 0.5 to land at 2688 x 1536. Save as PNG, crop and convert outside ComfyUI or with the Image Crop and Save Image (WebP) nodes from the Comfy core set. Total time on an RTX 4070: under two minutes including the upscale.

Optimising AI Images for WordPress Performance

AI-generated images are large files. A raw 2688 x 1536 PNG from an upscaler is often 8 to 12 MB. Without optimisation it will slow your site, hurt Largest Contentful Paint and eat your hosting storage. Here is how to handle it.

Convert to WebP or AVIF

WebP images are 25 to 35 percent smaller than JPEG at the same visual quality, and AVIF is smaller again, usually 40 to 50 percent under JPEG. WordPress has supported WebP uploads since 5.8 and AVIF since 6.5, and WordPress 6.7 added an option to convert uploads automatically where the server’s image library allows it. Most optimisation plugins handle the conversion and serve the right format to each browser.

Use Responsive Images

WordPress generates multiple sizes (thumbnail, medium, large, and 1536 and 2048 px variants) and serves the right one via the srcset attribute. Upload your AI image at the largest size you display and WordPress handles the rest. Do not disable the intermediate sizes to “save space”; mobile visitors will end up downloading the desktop file.

Recommended Optimisation Plugins

All four below were checked on WordPress.org in August 2026 and are actively maintained.

• ShortPixel Image Optimizer (v6.5.5): automatic compression with WebP and AVIF conversion on upload, bulk optimisation of the existing library, and a credit-based price that suits sites with bursts of uploads.
• Imagify (v2.3.3): three compression levels, WebP and AVIF output, and a clean interface. Made by the WP Rocket team, so it pairs well with that cache plugin.
• Smush (v4.3.0): free tier covers lossless compression and lazy loading; Pro adds WebP, a CDN and larger file limits. The highest-rated of the four on WordPress.org.
• EWWW Image Optimizer (v8.7.6): compresses on your own server without sending images to a third party, which matters for client work with privacy requirements. Optional paid API for stronger compression.

If images are only part of a bigger speed problem, our WordPress performance optimisation service covers caching, database cleanup and Core Web Vitals as a whole.

AI Image Tools Beyond Stable Diffusion

Stable Diffusion is not the only option, and in 2026 the open-weight lead has moved to FLUX. Here is how other popular AI graphic design tools compare for WordPress image generation:

• FLUX.2 (Black Forest Labs): open-weight dev and klein variants, up to 4 MP output at any aspect ratio, strong text rendering, multi-reference editing. The current default for self-hosted work if you have a 24 GB GPU. Hosted pro, flex and max tiers are available through the BFL API.
• Stable Diffusion 3.5: Large and Medium variants, 1 MP native, free under the Stability Community License for individuals and companies under $1M annual revenue. Good prompt adherence, smaller ecosystem than SDXL.
• GPT Image (ChatGPT / OpenAI API): generates at 1024 x 1024, 1536 x 1024 or 1024 x 1536. Excellent at following instructions and rendering text, no local control, per-image pricing.
• Midjourney: default 1024 x 1024 with upscales to 2048 and 4096, and a wide choice of aspect ratios via the --ar flag. Still the strongest for editorial and artistic styles.
• Adobe Firefly: built into Photoshop for generative fill and expand, trained on licensed content, so the safest choice for commercial client work.

The size rules are the same for all of them: find the native resolution, generate there, upscale and crop. Only the numbers change.

Common Mistakes to Avoid

• Generating at non-native resolution. Causes doubled subjects, artefacts and quality loss. Use the model’s trained bucket and upscale.
• Upscaling with too much denoise. Anything above 0.4 lets the upscaler redraw faces and hands. Keep it at 0.2 to 0.35 for photographic work.
• Uploading without compression. A single unoptimised AI image can be 5 to 12 MB. Convert to WebP or AVIF and compress before it touches the media library.
• Ignoring aspect ratio. Themes crop featured images to a fixed ratio. Generate close to that ratio so the crop does not cut off the subject.
• Forgetting text space. If a title or button will sit on the image, prompt for negative space on that side. Fixing it later means inpainting.
• Skipping alt text. AI images need descriptive alt text for accessibility and SEO. Describe what is in the image, not the prompt you used.
• Ignoring the licence. FLUX.1 and FLUX.2 dev weights are non-commercial, even though outputs can be used commercially. SD 3.5 is free below $1M revenue. Check before using a model for client deliverables.

Troubleshooting Size Problems

Two heads or repeated objects. Your canvas is too large for the model. Drop back to the native bucket and upscale instead.

Blurry or plasticky faces. Canvas is too small, or you upscaled with a plain resize. Use a model-based upscaler and a face-detailer pass.

Out of memory at 1024 x 1024. Enable the low-VRAM or tiled VAE options, switch to an fp8 or 4-bit quantised checkpoint, or use SDXL instead of FLUX.2 on cards under 12 GB.

“Dimensions must be multiples of 16”. SD 3.5 and FLUX need width and height divisible by 16. Round 1200 x 628 to 1200 x 624 or, better, generate at 1344 x 768 and crop.

WordPress shows a smaller image than I uploaded. The 2560 px big-image threshold scaled it down. That is fine for the web; upload at 2x your display size, not more.

Where to Start

If you are on a consumer GPU and want one setup that covers blog images, product shots and social cards, run SDXL or FLUX.1 dev in ComfyUI at 1024 x 1024 or 1344 x 768, upscale 2x with 4x-UltraSharp, crop in your editor, and push everything through ShortPixel or Imagify on upload. That pipeline covers every size in the WordPress table above and keeps pages fast.

If you have a 24 GB card and need readable text in images, move to FLUX.2 dev and generate hero banners at 2 to 4 MP directly, skipping the upscale step. And if you are producing images at volume for a community or store and would rather have the pipeline built and automated, our team does that as part of custom WordPress development.

Frequently asked questions

What is the best image size for Stable Diffusion?

512 x 512 for SD 1.5, 1024 x 1024 for SDXL and SD 3.5, and 1 to 2 megapixels for FLUX.1 (up to 4 MP on FLUX.2). Use the aspect-ratio buckets in the table above for non-square work and upscale afterwards.

Can Stable Diffusion generate 1920 x 1080 directly?

FLUX.2 can. SDXL and SD 3.5 will usually produce repeated subjects at 2 MP. Generate at 1344 x 768 or 1536 x 864 and upscale to 1920 x 1080 instead.

What size should a Stable Diffusion image be for a WordPress featured image?

Crop the final file to 1200 x 628 px. Generate at 1344 x 768, upscale 2x, then crop. Export as WebP under 150 KB.

Does a bigger image mean better quality?

Not from the base model. Quality comes from generating at the trained resolution and upscaling with a dedicated upscaler. A larger canvas only adds pixels the model does not know how to fill.

Is SDXL or FLUX better for WordPress images?

FLUX gives better text rendering and prompt adherence and handles more sizes natively. SDXL is lighter, runs on 8 GB cards and has a far larger library of LoRAs and fine-tunes. For most blog and product imagery either works; for banners with text, use FLUX.

---

High-Traffic Website Hosting: Kinsta Alternatives

Why is WordPress Free

SEO for Education Websites