# Stable Diffusion upscaler

> Upscale Stable Diffusion images with Real-ESRGAN, with nothing to install. The model runs in your browser, so nothing is uploaded.

Source: https://getoxtro.com/stable-diffusion-upscaler

This is an interactive tool. Use it at https://getoxtro.com/stable-diffusion-upscaler; it runs in the browser.

## What size are Stable Diffusion images?

SD 1.5 was trained at 512 x 512, and SDXL at 1024 x 1024, which Stability AI calls its native size. Most people generate at or near those sizes, because going much larger in a single pass tends to cause repeated subjects and other mistakes. So the usual way to get a large image is to generate small and upscale.

| Starting size | At 2x | At 4x |
| --- | --- | --- |
| 512 x 512 | 1024 x 1024 | 2048 x 2048 |
| 768 x 768 | 1536 x 1536 | 3072 x 3072 |
| 1024 x 1024 | 2048 x 2048 | 4096 x 4096 |

## How this compares with your local setup

If you run Automatic1111, Forge or ComfyUI, you may already have Real-ESRGAN and larger models. This page is for when you're on another computer, when you don't want to set up a workflow, or when you want a quick result to share.

It uses the compact Real-ESRGAN models: **Photo** is `realesr-general-x4v3` and **Art** is `realesr-animevideov3`. They were chosen because they're small enough to download in a browser. Heavier models such as `RealESRGAN_x4plus` can produce more detail, but they are far larger.

It is a pure upscaler. Unlike Hires. fix or tiled img2img, it doesn't run diffusion, so it won't invent new objects or follow your prompt.

## Frequently asked questions

### Which mode should I use for anime checkpoints?

Choose **Art**. It's built for line art and flat colour. Use **Photo** for realistic checkpoints and general scenes.

### Does it keep my generation settings?

No. The file you download is new, so metadata such as your prompt and seed isn't carried across. Keep the original if you need them.

### Can I run this on a batch of images?

Not at the moment. It takes one image at a time. Your local setup is the better choice for batches.