stable-diffusion

Stable Diffusion

Stable Diffusion is an open latent text-to-image diffusion architecture offering versatile fine-tuning (LoRA, ControlNet), local inference, and precise prompt typography adherence.

When to Use

• Open-Source Generative Image Synthesis: Generating photorealistic or stylized artwork using SDXL and Flux with Hugging Face diffusers.
• Precise Spatial Conditioning with ControlNet: Guiding image composition with depth maps, canny edges, and human poses.
• Inpainting, Outpainting & Image-to-Image: Modifying specific masked regions while preserving surrounding image context.
• Custom Style Adaptation with LoRA: Applying fine-tuned community style and character adapters.

Quick Start

import torch
from diffusers import StableDiffusionXLPipeline

# Load SDXL pipeline with FP16 precision
pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    variant="fp16"
).to("cuda")

prompt = "A high-tech digital laboratory, neon blue accents, volumetric lighting, photorealistic"
image = pipe(prompt=prompt, num_inference_steps=30).images[0]
image.save("output.png")

Core Concepts

Generating Images with SDXL & Hugging Face Diffusers

High-resolution generative image pipeline with Euler Ancestral scheduler:

import torch
from diffusers import AutoPipelineForText2Image, EulerAncestralDiscreteScheduler

pipeline = AutoPipelineForText2Image.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    variant="fp16",
    use_safetensors=True
).to("cuda")

# Configure fast Euler Ancestral scheduler
pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(pipeline.scheduler.config)

# Enable memory optimizations
pipeline.enable_vae_tiling()
pipeline.enable_xformers_memory_efficient_attention()

prompt = "cinematic macro photography of a crystal microchip illuminated by blue laser beams, 8k resolution, studio lighting"
negative_prompt = "low resolution, blurry, distorted, watermark, signature, artifacts"

image = pipeline(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=1024,
    height=1024,
    num_inference_steps=30,
    guidance_scale=7.0
).images[0]

image.save("crystal_chip.png")