Making pictures

The composer is the Composer side of the floating window. It's a form with one button: pick a model, describe what you want, press Generate.

Create, or Edit#

The composer works in one of two ways, and it decides which from what you're doing, not from a switch:

• Create — nothing to start from. It makes a new picture from your words.
• Edit — there's a picture to start from: the one open in the editor, a reference you've added, a selection you've drawn, or the AI inpaint brush in your hand. It changes that picture.

The big button follows: Generate on the Create side, Edit on the Edit side. Two small icons at the top — Text to image and Image to image — let you pin a side for now; the moment the work itself changes (you draw a selection, pick up a tool, add a picture), the composer goes back to reading it.

+ Editor puts the picture open in the editor into the composer as the thing to change. + Add brings in a picture of your own — a file, or a paste — as a reference. Video models also take a short clip or a sound file. Pictures are resized on the way in to fit the provider's limits; clips and sound must already be under 5 MB, and the composer says by how much one is over.

Picking a model#

The model picker lists every model your connected providers offer, pictures and video alike. It starts Cheapest first; a provider chip narrows it to Only <provider>, and a provider you haven't connected says no key yet. A model with several modes shows them as a row under the picker — Text to image · Image to image, for example.

Adding a model. ModelsLab hosts thousands of community checkpoints (SD 1.5 and SDXL). Add a model from ModelsLab opens a search — by name, or paste the model's ModelsLab link — and Add puts it in your list, where it behaves like any other model. Remove it from the list with a second click (Click again to remove it); its past renders stay.

The prompt#

What you type depends on the job, and the box tells you — A raven over a winter fjord, low sun, painted in oils… for a new picture, What to change — "make it night, the street lamp on"… for an edit, What fills the mask… for an inpaint. Remove background and Upscale need no words at all.

Two tools sit beside the prompt:

• Prompt history (Alt+↑ / Alt+↓) — every prompt you've sent, searchable, with Reuse prompt and model to bring one back. It keeps them across restarts.
• Prompt helper — a chat that knows each model's best practices. Ask it to write or fix a prompt for the model you've picked.

Settings#

Below the prompt, two folds:

• Negative prompt — What to keep out, on models that take one.
• Settings — Ratio, Size, a video's Length, a Seed (blank means random, so the same seed and prompt can make the same picture again), and whatever else this model offers. Those model-specific knobs come from the model itself, so they differ from one to the next.

A closed fold still shows a summary of what will be sent. Closed doesn't mean cleared.

LoRA appears on models that take them: small add-ons that steer a style or subject. Search, turn on up to three, and set each one's strength. Most need their trigger word in the prompt — it's on the LoRA's page. LoRA picks belong to the model's base (SD 1.5 or SDXL), and they're not kept after a restart.

The composer remembers the rest — model, prompt, settings, references — across restarts.

Where the render lands#

Press Generate or Edit. A card in the corner of the canvas shows the model and how long it's been running; most pictures take seconds, video longer. You can keep working meanwhile.

It lands in the editor. There's no results grid to pick from:

• Create, with a picture open and New layer on: a new top layer on that picture. With New layer off: its own picture, in a new tab.
• Edit, with New layer on: a new layer. Off: painted into the selected layer.
• Inpaint: only inside the selection.

Fit the picture to it decides who adapts: on, the canvas takes the render's size; off, the render is fitted to the picture.

Every render is also kept on the Semantix server with your renders. If one fails, the reason shows on the canvas or in the composer.

Where it is now#