https://github.com/runpod-workers/worker-sdxl
RunPod worker for Stable Diffusion XL
https://github.com/runpod-workers/worker-sdxl
Last synced: 12 months ago
JSON representation
RunPod worker for Stable Diffusion XL
- Host: GitHub
- URL: https://github.com/runpod-workers/worker-sdxl
- Owner: runpod-workers
- License: mit
- Created: 2023-08-08T08:21:08.000Z (almost 3 years ago)
- Default Branch: main
- Last Pushed: 2025-06-13T16:33:56.000Z (12 months ago)
- Last Synced: 2025-06-13T17:38:14.859Z (12 months ago)
- Language: Python
- Homepage:
- Size: 3.2 MB
- Stars: 32
- Watchers: 3
- Forks: 60
- Open Issues: 3
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
---
Run [Stable Diffusion XL](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) as a serverless endpoint to generate images.
---
[](https://www.runpod.io/console/hub/runpod-workers/worker-sdxl)
---
## Usage
The worker accepts the following input parameters:
| Parameter | Type | Default | Required | Description |
| :------------------------ | :------ | :------- | :-------- | :------------------------------------------------------------------------------------------------------------------ |
| `prompt` | `str` | `None` | **Yes\*** | The main text prompt describing the desired image. |
| `negative_prompt` | `str` | `None` | No | Text prompt specifying concepts to exclude from the image |
| `height` | `int` | `1024` | No | The height of the generated image in pixels |
| `width` | `int` | `1024` | No | The width of the generated image in pixels |
| `seed` | `int` | `None` | No | Random seed for reproducibility. If `None`, a random seed is generated |
| `scheduler` | `str` | `'DDIM'` | No | The noise scheduler to use. Options include `PNDM`, `KLMS`, `DDIM`, `K_EULER`, `DPMSolverMultistep` |
| `num_inference_steps` | `int` | `25` | No | Number of denoising steps for the base model |
| `refiner_inference_steps` | `int` | `50` | No | Number of denoising steps for the refiner model |
| `guidance_scale` | `float` | `7.5` | No | Classifier-Free Guidance scale. Higher values lead to images closer to the prompt, lower values more creative |
| `strength` | `float` | `0.3` | No | The strength of the noise added when using an `image_url` for image-to-image or refinement |
| `image_url` | `str` | `None` | No | URL of an initial image to use for image-to-image generation (runs only refiner). If `None`, performs text-to-image |
| `num_images` | `int` | `1` | No | Number of images to generate per prompt (Constraint: must be 1 or 2) |
| `high_noise_frac` | `float` | `None` | No | Fraction of denoising steps performed by the base model (e.g., 0.8 for 80%). `denoising_end` for base |
> [!NOTE]
> `prompt` is required unless `image_url` is provided
### Example Request
```json
{
"input": {
"prompt": "A majestic steampunk dragon soaring through a cloudy sky, intricate clockwork details, golden hour lighting, highly detailed",
"negative_prompt": "blurry, low quality, deformed, ugly, text, watermark, signature",
"height": 1024,
"width": 1024,
"num_inference_steps": 25,
"refiner_inference_steps": 50,
"guidance_scale": 7.5,
"strength": 0.3,
"high_noise_frac": 0.8,
"seed": 42,
"scheduler": "K_EULER",
"num_images": 1
}
}
```
which is producing an output like this:
```json
{
"delayTime": 11449,
"executionTime": 6120,
"id": "447f10b8-c745-4c3b-8fad-b1d4ebb7a65b-e1",
"output": {
"image_url": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAABAAAAAQACAIAAADwf7zU...",
"images": [
"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAABAAAAAQACAIAAADwf7zU..."
],
"seed": 42
},
"status": "COMPLETED",
"workerId": "462u6mrq9s28h6"
}
```
and when you convert the base64-encoded image into an actual image, it looks like this:
