Learn Stable Diffusion and ComfyUI
Open models you run yourself: SDXL, SD 3.5 and Flux in ComfyUI or the web UI. Checkpoints, LoRAs, samplers and the metadata in your files.
What Stable Diffusion & ComfyUI is good at
- Full control: checkpoints, LoRAs, ControlNet, samplers
- Running locally, no credits and no upload
- Repeatable results from seed and settings
- Prompt and workflow metadata saved inside the PNG
Quickstart
- Start with a known checkpointSDXL or SD 3.5 with the default sampler. Change one thing at a time.
- Fix the seedA fixed seed makes every other setting's effect visible.
- Add a LoRA at low weightStyle and character LoRAs at 0.6 to 0.8 before going higher.
- Keep the PNGThe web UI and ComfyUI write the prompt, seed and workflow into the file. Never flatten it.
Settings cheat-sheet
| Setting | What it does | Start with |
|---|---|---|
| Checkpoint | The base model | SDXL or SD 3.5 |
| Sampler / steps | How the image is denoised | Recommended sampler, 20 to 30 steps |
| CFG scale | Prompt adherence | 5 to 7 |
| Seed | Repeatable noise | Fixed while tuning |
| LoRA weight | Strength of an add-on | 0.6 to 0.8 |
| Resolution | Native size of the model first | 1024 px for SDXL |
Stable Diffusion & ComfyUI guides
Guides for Stable Diffusion & ComfyUI are on their way. The quickstart and cheat-sheet above are kept current in the meantime.
The five most-asked questions
What is the difference between ComfyUI and the web UI?
ComfyUI is node-based: you build the pipeline as a graph, which is powerful and shareable. The Stable Diffusion web UI (A1111 and forks) is form-based and quicker to start with.
What are checkpoints and LoRAs?
A checkpoint is the full model; a LoRA is a small add-on that teaches a style, character or object. LoRAs are applied with a weight and can be combined.
Which sampler and steps should I use?
Start with the model's recommended sampler and 20 to 30 steps, then adjust. More steps are not always better; compare at a fixed seed.
Where is my prompt saved?
Both tools embed the prompt, negative prompt, seed, sampler and (in ComfyUI) the whole workflow in the PNG's metadata. Keep the original file and you keep the recipe.
Can I use the images commercially?
It depends on the licence of each checkpoint and LoRA you used; Stability's models have their own community and enterprise terms. Read them and record which models made each image.