Stable Diffusion Cheat Sheet
A quick reference for Stable Diffusion: the key settings (CFG, steps, sampler, seed), prompt and negative-prompt tips, and the models and tools that shape your images.
Key settings
| Setting | Description | Example / tip |
|---|---|---|
| What you want to see - subject, style, lighting and detail. | portrait of a fox, oil painting, soft light | |
| What to avoid - the fastest way to clean up common artifacts. | blurry, extra fingers, text, watermark | |
| How strictly it follows the prompt. Higher is more literal, lower is more creative. 5-9 is a good range. | CFG 7 | |
| How many refinement steps. More can add detail but with diminishing returns. | 20-30 steps | |
| The algorithm that denoises the image; different samplers give slightly different looks. | DPM++ 2M Karras | |
| Fix the seed to reproduce an image or make controlled variations. | Reuse a seed, tweak the prompt |
Go further
Ready to master it for real?
This reference is a great start. Go from the basics to pro workflows in our in-depth Leonardo AI course.
Image to image & control
| Setting | Description | Example / tip |
|---|---|---|
| In img2img, how much the source image is changed. Low keeps it close, high reinvents it. | 0.4 = subtle, 0.75 = big change | |
| Repaint only a masked area to fix or replace part of an image. | Mask the hands, regenerate | |
| Guide generation with a reference for pose, depth, edges or composition. | Match a pose from a photo | |
| Increase resolution and add detail after generating. | Hires fix or an upscaler |
Models & extras
| Setting | Description | Example / tip |
|---|---|---|
| The main model sets the overall style and quality (e.g. SDXL). | Pick a photoreal or anime checkpoint | |
| Small add-on models that inject a specific style, character or concept. | Apply a style LoRA at 0.6 weight | |
| Affects final color and sharpness; the right VAE fixes washed-out output. | Use the recommended VAE for your model | |
| Emphasize or de-emphasize words to steer the result. | (golden hour:1.3) |
Frequently asked questions
What does CFG scale do?
It controls how closely Stable Diffusion follows your prompt. Lower values give the model more creative freedom, higher values stick more literally to the words but can look over-cooked. Somewhere around 6-8 is a reliable starting range.
How do I fix bad hands and artifacts?
Add them to the negative prompt (for example 'extra fingers, deformed hands'), use a model known for anatomy, lower the denoising strength in img2img, and inpaint the problem area to regenerate just that part.
What's the difference between a checkpoint, a LoRA and ControlNet?
A checkpoint is the full base model that sets the overall look. A LoRA is a small add-on that injects a specific style or character on top. ControlNet steers the composition using a reference like a pose or depth map. They stack together.
Is Stable Diffusion free?
The models are open and free to run yourself locally or through various web interfaces, though hosted services and heavier GPUs may cost money.
More cheat sheets
Just getting started? Try a free course
Beginner-friendly courses in Blender, Photoshop, Figma, Canva and Claude - free.
Go from reference to real skills
Cheat sheets are great for quick lookups. Our in-depth courses take you from the fundamentals to professional-level mastery.
Master Leonardo AI with the full course