Stable Diffusion Image Generation

Instructions

Generate images using the Stable Diffusion backend of the ITI image generation service at http://localhost:7860 (native) or http://stable-diffusion:7860 (Docker network).

Note: The image generation service now supports three backends: sd, flux-local, and flux-api. For FLUX-based generation (better photorealism, text rendering, hex color control), see the flux-image-generation skill. This skill covers the legacy Stable Diffusion backend.

Text-to-image generation:

curl -X POST http://localhost:7860/api/txt2img \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "a professional product photo of a laptop on a minimalist desk, soft lighting, 4k",
    "negative_prompt": "blurry, low quality, distorted, watermark",
    "backend": "sd",
    "width": 768,
    "height": 512,
    "num_inference_steps": 50,
    "guidance_scale": 7.5,
    "seed": 42
  }'

Image-to-image transformation:

curl -X POST "http://localhost:7860/api/img2img?prompt=oil+painting+style&strength=0.7&backend=sd" \
  -F "image=@input.png"

Prompt engineering for Stable Diffusion:

• Front-load the subject: “A golden retriever puppy, sitting in a garden”
• Add style qualifiers: “digital art”, “oil painting”, “photorealistic”, “watercolor”, “3D render”
• Add quality boosters: “highly detailed”, “4k”, “sharp focus”, “professional lighting”
• Use negative prompts to exclude: “blurry, low quality, distorted, bad anatomy, watermark, text”
• Weights work in some models: “(subject:1.3)” increases emphasis

Parameter tuning guide:

Model selection:

When to choose SD over FLUX:

• You need community fine-tunes or LoRAs (SD has the largest ecosystem)
• You’re working with an existing SD-based workflow
• VRAM is limited (~5 GB vs ~13 GB for FLUX Klein)
• Specific SD model fine-tunes match your use case

Resource awareness:

• Docker CPU mode: ~120s per 512×512 image, ~4 GB RAM
• Native MPS (M3 Max): ~8-12s per 512×512 image, ~5 GB RAM
• For batch generation, prefer native mode to avoid Docker CPU overhead
• First generation after startup is slower due to model loading (~30-60s additional)

Related skills:

• flux-image-generation — FLUX models for superior photorealism, text rendering, and color control
• flux-operations — FLUX backend deployment, BFL API credentials, Klein local setup
• stable-diffusion-operations — SD service deployment and management