Stable Diffusion 3.5 Large Turbo Worker Template

A specialized worker template for building custom RunPod Endpoint API workers utilizing the Stable Diffusion 3.5 Large Turbo model. This implementation supports various aspect ratios, guidance scales, and inference steps customization.

Docker Container

The ready-to-use Docker container is available on Docker Hub:

thehunter911/stbldiff3.5-large-turbo-runpod-serverless

which can be used to deploy onto Runpod Serverless endpoint directly.

You can also pull it using:

docker pull thehunter911/stbldiff3.5-large-turbo-runpod-serverless

Input Parameters

Parameter	Type	Default	Description
prompt	string	required	The text description of the image you want to generate
aspect_ratio	string	"1:1"	Image aspect ratio. See supported values below
guidance_scale	float	0.0	How closely the model follows the prompt
num_inference_steps	integer	1	Number of denoising steps

Supported Aspect Ratios

• "16:9" - Widescreen (1024×576)
• "1:1" - Square (1024×1024)
• "21:9" - Ultra-wide (1024×439)
• "2:3" - Portrait (683×1024)
• "3:2" - Landscape (1024×683)
• "4:5" - Portrait (819×1024)
• "5:4" - Landscape (1024×819)
• "9:16" - Vertical/Mobile (576×1024)
• "9:21" - Vertical ultra-wide (439×1024)

Example Input

{
    "input": {
        "prompt": "An image of a cat with a hat on",
        "aspect_ratio": "16:9",
        "guidance_scale": 0.0,
        "num_inference_steps": 1
    }
}

Example Output

The output is a base64 encoded string of the generated image. Example:

data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAA...

Viewing the Output

To view the generated image, you can decode the base64 string using Python:

import base64
from PIL import Image
import io

# Replace 'base64_string' with your actual base64 string
base64_string = "iVBORw0KGgoAAAANSUhEUgAA..."
image_data = base64.b64decode(base64_string)
image = Image.open(io.BytesIO(image_data))
image.show()

Performance Notes

This implementation uses torch.float16 with fp16 variant for optimal inference performance on consumer GPUs. This configuration provides:

• Faster inference speed
• Lower memory usage
• Better hardware compatibility
• Optimal for generation tasks