ThinkExponential AILab — Stable Diffusion Usage Guide

This guide covers how to generate images with AUTOMATIC1111 Stable Diffusion WebUI running locally in Docker.

Prerequisite: Complete environment-setup.md and start the Stable Diffusion stack before following this guide.

Start the stack
Download a model
WebUI walkthrough
Text-to-image (txt2img)
Image-to-image (img2img)
REST API examples
Python client examples
Per-model guide and examples
Prompt engineering guide
Parameter reference
Managing models

1. Start the stack

From the repository root:

# NVIDIA GPU
docker compose -f stable-diffusion/docker-compose.nvidia.yml up -d

# AMD GPU
docker compose -f stable-diffusion/docker-compose.amd.yml up -d

# CPU only (slow — minutes per image)
docker compose -f stable-diffusion/docker-compose.cpu.yml up -d

The first startup takes a few minutes while it downloads the base image layers. Open http://localhost:7860 to access the WebUI.

2. Download a model

Models are stored in the sd_models Docker volume and persist across restarts.

Option A — download from Hugging Face via the WebUI

Open http://localhost:7860
Go to the Civitai Helper or Model Downloader extension (if installed), or use the terminal method below.

Option B — download directly inside the container

# Open a shell inside the container
docker exec -it stable-diffusion-webui bash

# Navigate to the models folder
cd /app/stable-diffusion-webui/models/Stable-diffusion/

# Download SD 2.1 (requires Hugging Face login for some models)
wget -q "https://huggingface.co/stabilityai/stable-diffusion-2-1/resolve/main/v2-1_768-ema-pruned.safetensors"

# Exit the container shell
exit

Option C — copy a local file into the container

docker cp ~/Downloads/my-model.safetensors \
  stable-diffusion-webui:/app/stable-diffusion-webui/models/Stable-diffusion/

After copying, click 🔄 Refresh next to the model dropdown in the WebUI.

Model download links

Model	File	Size
SD 2.1	v2-1_768-ema-pruned.safetensors	5.2 GB
SDXL base	sd_xl_base_1.0.safetensors	6.9 GB
SDXL refiner	sd_xl_refiner_1.0.safetensors	6.1 GB
FLUX.1 [schnell]	See Hugging Face	~24 GB
FLUX.1 [dev]	See Hugging Face	~24 GB (gated)
SD 3.5 Large	See Hugging Face	~16 GB (gated)

Gated models require a Hugging Face account and model access request. Log in with huggingface-cli login inside the container and use the access token.

3. WebUI walkthrough

Open http://localhost:7860.

Main tabs

Tab	Purpose
txt2img	Generate images from a text prompt
img2img	Transform or vary an existing image
Extras	Upscale, face restoration
PNG Info	Read generation parameters from a saved image
Settings	Configure defaults, extensions, and server options

Key UI controls

Checkpoint (top-left) — select which model file to use; click 🔄 to refresh after adding new files
Prompt — positive prompt (what to include)
Negative prompt — what to exclude
Sampling method — the diffusion algorithm (DPM++ 2M Karras is a reliable default)
Sampling steps — more steps = higher quality but slower (20–30 is a good range)
Width / Height — output image resolution
CFG Scale — how closely to follow the prompt (7–12 is typical)
Seed — -1 for random; fix a seed to reproduce an image exactly

4. Text-to-image (txt2img)

Basic generation

Select a model from the Checkpoint dropdown.
Enter a positive prompt in the top text box.
(Optional) Enter a negative prompt.
Click Generate.

Example prompts

Landscape photography:

Prompt:
  a misty mountain valley at sunrise, golden hour light, pine forest,
  photorealistic, 8k, shot on Canon EOS R5, depth of field

Negative prompt:
  cartoon, painting, blurry, low quality, watermark, text

Portrait:

Prompt:
  portrait of a woman in her 30s, soft studio lighting, bokeh background,
  professional headshot, high resolution, sharp focus

Negative prompt:
  deformed, ugly, bad anatomy, extra limbs, watermark, signature

Concept art:

Prompt:
  futuristic city at night, neon reflections on wet street, flying cars,
  cyberpunk aesthetic, detailed, artstation, by Greg Rutkowski

Negative prompt:
  blurry, low resolution, oversaturated, distorted

5. Image-to-image (img2img)

Img2img lets you start from an existing image and guide the generation with a prompt.

Steps

Click the img2img tab.
Upload a source image by dragging it into the canvas or clicking the upload icon.
Enter a prompt describing what you want the output to look like.
Adjust Denoising strength:
- 0.3–0.5 — subtle variation, stays close to original
- 0.6–0.8 — significant change
- 0.9–1.0 — almost entirely new image
Click Generate.

Common img2img use cases

Style transfer — change the art style while keeping composition
Inpainting — fill in or replace part of an image (use the Inpaint sub-tab)
Upscaling with detail — use img2img with low denoising at 2× resolution
Sketch to image — turn a rough sketch into a rendered image

6. REST API examples

The WebUI exposes a REST API at http://localhost:7860/sdapi/v1/. The full Swagger UI is available at http://localhost:7860/docs.

txt2img

curl -s -X POST http://localhost:7860/sdapi/v1/txt2img \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "a cozy cabin in a snowy forest, warm light from windows, evening, photorealistic",
    "negative_prompt": "blurry, low quality, cartoon, painting",
    "steps": 25,
    "cfg_scale": 7.5,
    "width": 512,
    "height": 512,
    "sampler_name": "DPM++ 2M Karras",
    "seed": -1
  }' | python3 -c "
import sys, json, base64
data = json.load(sys.stdin)
img = base64.b64decode(data['images'][0])
open('output.png', 'wb').write(img)
print('Saved output.png')
"

img2img

# Encode the source image to base64
BASE64_IMAGE=$(base64 -w 0 input.png)

curl -s -X POST http://localhost:7860/sdapi/v1/img2img \
  -H "Content-Type: application/json" \
  -d "{
    \"init_images\": [\"${BASE64_IMAGE}\"],
    \"prompt\": \"oil painting style, impressionist, vibrant colors\",
    \"denoising_strength\": 0.65,
    \"steps\": 20,
    \"cfg_scale\": 7,
    \"width\": 512,
    \"height\": 512
  }" | python3 -c "
import sys, json, base64
data = json.load(sys.stdin)
img = base64.b64decode(data['images'][0])
open('output_img2img.png', 'wb').write(img)
print('Saved output_img2img.png')
"

Get available models

curl http://localhost:7860/sdapi/v1/sd-models | python3 -m json.tool

Switch active model

curl -X POST http://localhost:7860/sdapi/v1/options \
  -H "Content-Type: application/json" \
  -d '{"sd_model_checkpoint": "v2-1_768-ema-pruned.safetensors"}'

Get current settings

curl http://localhost:7860/sdapi/v1/options | python3 -m json.tool

7. Python client examples

pip install requests pillow

Generate and save an image

import requests
import base64
from PIL import Image
from io import BytesIO

def txt2img(prompt: str, negative_prompt: str = "", steps: int = 25, **kwargs) -> Image.Image:
    payload = {
        "prompt": prompt,
        "negative_prompt": negative_prompt,
        "steps": steps,
        "cfg_scale": kwargs.get("cfg_scale", 7.5),
        "width": kwargs.get("width", 512),
        "height": kwargs.get("height", 512),
        "sampler_name": kwargs.get("sampler_name", "DPM++ 2M Karras"),
        "seed": kwargs.get("seed", -1),
    }
    response = requests.post("http://localhost:7860/sdapi/v1/txt2img", json=payload)
    response.raise_for_status()
    image_data = base64.b64decode(response.json()["images"][0])
    return Image.open(BytesIO(image_data))


# Generate an image
image = txt2img(
    prompt="a serene Japanese garden with a koi pond, cherry blossoms, golden hour",
    negative_prompt="people, text, watermark, low quality",
    steps=30,
    cfg_scale=8,
    width=768,
    height=512,
)
image.save("japanese_garden.png")
print("Saved japanese_garden.png")

Batch generation

import requests, base64
from PIL import Image
from io import BytesIO

prompts = [
    "a red apple on a wooden table, photorealistic",
    "a blue sports car on a mountain road, cinematic",
    "an abstract painting with swirling galaxies, vibrant colors",
]

for i, prompt in enumerate(prompts):
    response = requests.post(
        "http://localhost:7860/sdapi/v1/txt2img",
        json={"prompt": prompt, "steps": 20, "width": 512, "height": 512},
    )
    image_data = base64.b64decode(response.json()["images"][0])
    image = Image.open(BytesIO(image_data))
    image.save(f"batch_{i}.png")
    print(f"Saved batch_{i}.png")

8. Per-model guide and examples

Stable Diffusion 2.1

SD 2.1 uses a 768×768 native resolution. Best for general photography, landscapes, and portraits with simple prompts.

Recommended settings:

Sampler: DPM++ 2M Karras
Steps: 20–30
CFG Scale: 7–9
Resolution: 768×768

Example:

Prompt:
  majestic eagle soaring above snow-capped mountains, dramatic clouds,
  golden sunlight, wildlife photography, Canon EF 500mm, sharp focus

Negative prompt:
  blurry, overexposed, low detail, cartoonish

Stable Diffusion XL (SDXL)

SDXL produces 1024×1024 outputs and can be chained with the refiner model for enhanced detail.

Recommended settings:

Sampler: DPM++ 2M Karras
Steps: 30–40
CFG Scale: 6–8
Resolution: 1024×1024

Switch to SDXL:

curl -X POST http://localhost:7860/sdapi/v1/options \
  -H "Content-Type: application/json" \
  -d '{"sd_model_checkpoint": "sd_xl_base_1.0.safetensors"}'

Example:

Prompt:
  hyperrealistic portrait of an elderly sailor with weathered skin,
  piercing blue eyes, white beard, dramatic ocean background,
  studio lighting, 8k, photorealistic, Rembrandt lighting

Negative prompt:
  cartoon, painting, unrealistic, bad anatomy, extra fingers, watermark

Using base + refiner (two-pass):

import requests, base64
from PIL import Image
from io import BytesIO

# Pass 1: base model generates a latent
base_response = requests.post("http://localhost:7860/sdapi/v1/txt2img", json={
    "prompt": "a futuristic cityscape at dusk, neon lights reflecting on glass towers",
    "steps": 40,
    "cfg_scale": 7,
    "width": 1024,
    "height": 1024,
    "denoising_strength": 0.8,
}).json()

# Save base output
base_image_bytes = base64.b64decode(base_response["images"][0])
Image.open(BytesIO(base_image_bytes)).save("sdxl_base.png")

# Pass 2: refiner polishes the output
requests.post("http://localhost:7860/sdapi/v1/options", json={
    "sd_model_checkpoint": "sd_xl_refiner_1.0.safetensors"
})
refiner_response = requests.post("http://localhost:7860/sdapi/v1/img2img", json={
    "init_images": [base_response["images"][0]],
    "prompt": "a futuristic cityscape at dusk, neon lights reflecting on glass towers",
    "denoising_strength": 0.3,
    "steps": 20,
    "cfg_scale": 7,
    "width": 1024,
    "height": 1024,
}).json()

Image.open(BytesIO(base64.b64decode(refiner_response["images"][0]))).save("sdxl_refined.png")
print("Saved sdxl_base.png and sdxl_refined.png")

Stable Diffusion 3.5

SD 3.5 is Stability AI's most recent release. It features significantly improved text rendering, prompt adherence, and composition. It requires at least 10 GB VRAM for the full model.

Recommended settings:

Sampler: Euler
Steps: 28
CFG Scale: 4.5
Resolution: 1024×1024

SD 3.5 uses a different architecture (MMDiT). Ensure the WebUI version supports it, or use a dedicated SD 3.5-compatible loader.

Example:

Prompt:
  A glass bottle containing a tiny universe with swirling galaxies inside,
  placed on an ancient wooden table, dramatic lighting, hyperrealistic macro photography

Negative prompt:
  (none needed — SD 3.5 follows positive prompts very well)

FLUX.1

FLUX.1 by Black Forest Labs achieves state-of-the-art image quality. It requires 16–24 GB VRAM for full performance. The schnell variant is faster; dev is higher quality.

Recommended settings:

Sampler: Euler
Steps: 4 (schnell) / 20–28 (dev)
CFG Scale: 1 (schnell) / 3.5 (dev)
Resolution: 1024×1024

FLUX.1 requires the sd-forge or ComfyUI frontend for best compatibility. AUTOMATIC1111 may require an extension.

Example:

Prompt (FLUX.1 schnell — 4 steps):
  a photorealistic image of a golden retriever puppy playing in autumn leaves,
  shallow depth of field, warm afternoon light

Negative prompt:
  (FLUX.1 does not use negative prompts — leave empty)

# FLUX.1 schnell — fast generation (4 steps)
curl -s -X POST http://localhost:7860/sdapi/v1/txt2img \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "a photorealistic image of a golden retriever puppy playing in autumn leaves, shallow depth of field",
    "steps": 4,
    "cfg_scale": 1,
    "width": 1024,
    "height": 1024,
    "sampler_name": "Euler"
  }' | python3 -c "
import sys, json, base64
data = json.load(sys.stdin)
open('flux_output.png', 'wb').write(base64.b64decode(data['images'][0]))
print('Saved flux_output.png')
"

9. Prompt engineering guide

Positive prompt structure

A well-structured prompt typically follows this pattern:

[subject], [setting/context], [style descriptors], [lighting], [camera/technical details], [quality boosters]

Example:

a lone wolf standing on a cliff, overlooking a misty valley,
oil painting style, dramatic, chiaroscuro lighting,
highly detailed, masterpiece, artstation

Effective quality boosters

photorealistic, hyperrealistic, 8k, ultra-detailed, sharp focus, 
masterpiece, best quality, professional, award-winning photography

Effective negative prompt terms

blurry, low quality, low resolution, bad anatomy, deformed hands, 
extra fingers, watermark, signature, text, username, artist name,
ugly, pixelated, overexposed, underexposed, grainy, noisy

Prompt weighting

Increase or decrease the influence of specific terms using parentheses:

# Increase weight (default multiplier: 1.1 per level)
(bright sunlight:1.4) a forest path

# Decrease weight
a forest path [rain:0.5]

Negative prompt for anatomy

When generating people, a detailed negative prompt helps:

bad anatomy, bad hands, extra fingers, missing fingers, fused fingers,
too many fingers, deformed, mutated, disfigured, poorly drawn face,
poorly drawn hands, extra limbs, missing limbs, floating limbs,
disconnected limbs, malformed limbs, ugly, duplicate, morbid

10. Parameter reference

Parameter	Typical range	Notes
Steps	20–40	More steps = better quality, slower. Diminishing returns above 40
CFG Scale	5–12	How strictly to follow the prompt. Higher = more literal but may lose realism
Denoising strength	0.3–0.9	img2img only. Lower = more similar to input image
Width / Height	512–1024	Stay near the model's native resolution. SDXL/FLUX: 1024. SD 2.1: 768
Seed	-1 or any int	-1 = random. Fix seed to reproduce images exactly
Batch size	1–4	Number of images per generation (uses more VRAM)
Batch count	1–8	Number of sequential generation runs

Samplers comparison

Sampler	Speed	Quality	Notes
Euler	Fast	Good	Reliable default, works well with FLUX
Euler a	Fast	Good	More variation per seed
DPM++ 2M Karras	Medium	Very good	Best general-purpose sampler for SD 1.x/2.x
DPM++ SDE Karras	Slow	Excellent	Highest detail, less predictable
DDIM	Medium	Good	Deterministic, useful for img2img
LMS Karras	Fast	Good	Fast convergence

11. Managing models

List installed models

curl http://localhost:7860/sdapi/v1/sd-models | python3 -c "
import sys, json
for m in json.load(sys.stdin):
    print(m['model_name'])
"

Check current model

curl http://localhost:7860/sdapi/v1/options | python3 -c "
import sys, json
opts = json.load(sys.stdin)
print('Current model:', opts.get('sd_model_checkpoint'))
"

Switch model via API

curl -X POST http://localhost:7860/sdapi/v1/options \
  -H "Content-Type: application/json" \
  -d '{"sd_model_checkpoint": "sd_xl_base_1.0.safetensors"}'

Remove a model (free disk space)

# Enter the container and delete the file
docker exec -it stable-diffusion-webui \
  rm /app/stable-diffusion-webui/models/Stable-diffusion/old-model.safetensors

Inspect volume disk usage

docker system df -v | grep sd_models

FilesExpand file tree

stable-diffusion-usage.md

Latest commit

History

stable-diffusion-usage.md

File metadata and controls

ThinkExponential AILab — Stable Diffusion Usage Guide

Table of contents

1. Start the stack

2. Download a model

Option A — download from Hugging Face via the WebUI

Option B — download directly inside the container

Option C — copy a local file into the container

Model download links

3. WebUI walkthrough

Main tabs

Key UI controls

4. Text-to-image (txt2img)

Basic generation

Example prompts

5. Image-to-image (img2img)

Steps

Common img2img use cases

6. REST API examples

txt2img

img2img

Get available models

Switch active model

Get current settings

7. Python client examples

Generate and save an image

Batch generation

8. Per-model guide and examples

Stable Diffusion 2.1

Stable Diffusion XL (SDXL)

Stable Diffusion 3.5

FLUX.1

9. Prompt engineering guide

Positive prompt structure

Effective quality boosters

Effective negative prompt terms

Prompt weighting

Negative prompt for anatomy

10. Parameter reference

Samplers comparison

11. Managing models

List installed models

Check current model

Switch model via API

Remove a model (free disk space)

Inspect volume disk usage