Generate
Type what you want and press Render. That is the whole minimum.
Plain description works better than a list of keywords. "A red fox in snow, soft winter light" beats "fox, snow, winter, 8k, masterpiece".
The row underneath offers Things to avoid (what not to draw), Improve my wording, and preset chips — Portrait photo, Full body, Scenery, Product, Illustration — which set sensible sizes and step counts for that kind of picture.
| Portrait 832×1216 | Full body, standing figures |
| Square 1024×1024 | Faces, product shots — usually the safest |
| Landscape 1216×832 | Scenes and environments |
These are SDXL's own native sizes. Asking for something far from them tends to produce duplicated limbs and stretched faces.
Quick look is a draft. Finished takes longer and resolves skin, fabric and edges. The button tells you roughly how long, measured on your machine rather than guessed.
- How much detail (steps) — more passes, more fine texture. Past a point it stops helping and only costs time.
- How closely it follows your words (CFG) — low lets it invent, high follows you literally and can go stiff. Around 6 is balanced.
- Get the same picture again (seed) — lock it to reproduce a render exactly, re-roll for a different take on the same prompt.
- Hand the last of it to a second model (refiner) — an SDXL-only extra pass for final polish.
Adapters you have trained appear on the right. A subject adapter does nothing until its trigger phrase is in the prompt, so switching one on shows the phrase, a + button that inserts it for you, and a warning while it is missing. A style adapter has no trigger and works the moment it is on.
See Library and Training a LoRA.
Several seconds to half a minute, because the model has to load into the graphics card. Every render after that is much faster. The estimate on the button accounts for both.