You can see the quality of the output below that was generated in under 3 seconds.

Use the endpoint URL and Python code to run OctoAI's fastest SDXL in your app.

import os
from octoai.clients.image_gen import Engine, ImageGenerator

if __name__ == "__main__":
    image_gen = ImageGenerator(token=os.environ.get("OCTOAI_TOKEN"))
    image_gen_response = image_gen.generate(
        engine=Engine.SDXL,
        prompt="A pug dog romping in a meadow of flowers on a sunny day",
        negative_prompt="Blurry photo, distortion, low-res, poor quality",
        width=1024,
        height=1024,
        num_images=1,
        sampler="DDIM",
        steps=30,
        cfg_scale=12,
        use_refiner=True,
        high_noise_frac=0.8,
        style_preset="base",
    )
    images = image_gen_response.images

    for i, image in enumerate(images):
        image.to_file(f"result{i}.jpg")

Fine-tuning included

Upload images to generate fine tuned stable diffusion weights using OctoAI’s LoRA fine-tuning service to align models to your brand or style.

Take your base models's creativity to infinity

Augment SD 1.5, and SDXL base models with customized LoRAs, VAEs, textual inversions, hypernetworks, refiners and Dreambooth checkpoints to unlock limitless creative styles and new aesthetic concepts.

Go beyond image generation

Mix and match variants of stable diffusion and image manipulation models such as ControNet, Segment Anything and CLIP to build sophisticated workflows to quickly achieve your creative goals.

Stable Diffusion on OctoAI features

Features	OctoAI
Load custom model weights via LoRAs, textual inversions, VAEs, etc	Yes via API
Mix and match image generation & enhancement models into workflows	Yes link to API
Time per SDXL image generated	0.6 sec for base model or 1.4 sec for custom checkpointed model
Time per SD1.5 image generated	3.8 sec for base model or 5.8 sec for custom checkpointed model

Stable Diffusion model history

Run time information

OctoAI brings you the fastest Stable Diffusion 1.5 and SDXL available today. Internal benchmarking for SDXL, shows consistently generated 1024x1024 pixel resolution 30-step images on SDXL 1.0 with a p95 time of 2.8 seconds.