Iris
Image Studio's own agent. She sees the canvas, works the editor with the same tools you do, renders when you ask, and remembers how you like to edit.
Iris is the agent who lives in Image Studio. She isn't the chat you code with, and she isn't the Studio agent: her whole job is the picture in front of you, and the edit you're making to it.
She's the Agent side of the floating window over the editor. Switch to Composer for the form; switch back to talk to her.
What she sees#
Every message you send her carries the editor as it stood when you sent it — the picture, its layers and which one is active, the selection, the tool in your hand — and a small picture of the canvas. So "this layer" and "the selection" mean what you mean. When she needs a closer look, she can look at one layer alone, the selected pixels, or a part of the picture at full size.
Seeing needs a model that can see pictures. On one that can't, she says so and works from the layers and your description.
Note
Each look is a vision call on your model, so it costs like one. She looks once per question, and again only after something has changed.
What she can do#
Explain, or do. Ask how and she answers with the steps, named the way the editor names them. Ask her to do it and she does, with the same tools you have: layers, masks, selections, adjustment layers, brush and eraser strokes, gradients, text, transforms, undo. Or ask her to set a tool up for you — "a soft red brush, 40 px" — and you paint.
She prefers edits that stay editable: a new layer over painting the original, an adjustment layer for light and colour, a mask instead of the eraser. Every change she makes is one undo step. Before something that loses work — delete, merge, flatten — she says what will go.
Render. She can make a picture from words, change the open picture with an AI edit, and make video clips from words, from a picture, or between two pictures. She renders only when you ask — no extra takes, no variations, no retries you didn't ask for — and says which model she started, and its price, before she spends anything. A render lands on its own a moment later as a new top layer; she's told when it lands and looks at it before saying it worked. A video clip doesn't land on the canvas.
Export. Ask for "a JPG for the web" and she writes it to ~/.semantix/media/images, the same as
Export….
What she remembers#
Iris keeps a memory of her own across sessions — how you edit, what you like — and uses it without being asked. It's her memory, not a project's. It follows the memory setting under Settings → Privacy.
Her model#
She has a model picker of her own, in her message box beside Send. Pick one that can see pictures; she does her best work when she can look.
What she won't do#
Nothing sexual involving anyone under 18 or anyone who looks it, and no sexual edits of a real person's photo. She'll say so plainly and offer what she can do instead.
The Prompt helper#
Image Studio has a second helper, for the composer rather than the canvas: the Prompt helper, from the button beside the prompt. It knows each model's best practices and writes prompts for the model you've picked. Iris works the picture; the Prompt helper works the form.