Box Director

Build the image box by box: everything lands where you drew it, a logo or a title is placed to the pixel, one area is redone without touching the rest, and what is frozen no longer moves. Online in the axiris.ai interface.

Box Director is the image direction tool from Axiris. You place boxes on the image, each with its position, its description and its role: an object, exact text, an image of your own, a shape. The model works inside those boxes; what comes out right is frozen and no longer moves; one area is redone with the brush without touching the rest; a composition is saved as a layout and placed on the next image. It is online in the axiris.ai interface, for accounts with access to generation.

What Box Director is

Box Director is the image direction tool from Axiris. In it, an image is a work surface. You place boxes on it, each with its position, its description and its role: an object, text, an image of your own, a shape. The model works inside those boxes. What it gets right, you freeze, and it no longer moves. What it misses, you redo, inside its box or in an area painted with the brush, without touching the rest.

What it changes

Elsewhere, an image is a draw: you run it again, you hope, and everything moves with every attempt. The product is right but the title has changed; the title is right but the character is no longer the same. A client who has approved an element can watch it vanish on the next run.

With Box Director, an image is built like a mock-up. Every element has its place before it exists. A logo, a packaging, a material come in as they are: placed, not redrawn. A title is set to the pixel, in the chosen typeface. A face, a garment come from one of your reference images. And when an area is redone, the rest has not changed: "the rest does not move" is not a figure of speech, it is what we check on the pixels.

A composition that works is saved as a layout and placed on the next image: a poster is built once, then declined into variants.

Where it is

Box Director is online, in the axiris.ai interface, for accounts with access to generation. Nothing to install: everything happens in the page. Nova sits next to the board: she helps you write the boxes, the scene, the background; you decide. The guide below describes the page, gesture by gesture.

The guide to the Box page

We no longer generate an image: we build it. You draw boxes on the image, you say what goes in each one, and everything lands where you drew it. What comes out right, you freeze: it no longer moves. The rest is redone around it.

[screenshot: the whole Box page: Nova on the left, the board in the centre, the settings on the right]

The page at a glance

On the left, Nova: the same conversation as in Advanced generation. She helps you write your boxes, your scene, your background.
In the centre, the board: your image and its boxes. The wheel (with Ctrl) zooms, the zoom level is at the bottom right, the render status at the top right.
On the right, the settings: the format, the references, the number of images, the scene (Summary) and the background (Background), then the board's toolbar.
At the bottom, the brief bar. Ctrl+P hides it: the board extends to the bottom of the screen. Ctrl+P brings it back.

1 · Drawing the boxes

Pick a tool in the toolbar, then drag a rectangle on the board. The tool then returns to ↖ (move).

Every box has a number (01, 02...) that never changes, even if you delete others.

Tell the model how, in the description. The box says where; the description can say how: small, in the bottom right corner, a large title across the top, half hidden behind the vase. The closer the description matches the box, the better it is followed.

The scene and the background. Summary: the whole image in one or two sentences. Background: the setting behind everything; leave it empty and the model invents it. Anything named only in the scene has no reserved place: give it a box, or write it as a sentence beginning with add (add a big dragon at right).

▶ Generate (or Generate at the bottom) starts the render. It appears under your boxes.

[screenshot: a board with three boxes and its render]

2 · Your reference images: `<Picture 2>`, `<Picture 3>`...

In References, drop an image: it becomes `<Picture 2>`, the next one `<Picture 3>`, up to 9. Name it in a box:

the woman from `<Picture 2>`, leaning against the car;
the car from `<Picture 3>`, three-quarter front view.

The box says where, your words say what is happening, the image says who or what. The model redraws it in the scene, in the requested pose and light: it does not paste it.

After a render, a box that calls a reference EDITS the displayed image: only those boxes change, the rest does not move by a single pixel. `<Picture 1>` is the displayed render itself. If you have changed the scene, the background, the format or a box without a reference, the page first recreates the image, then edits.

3 · ❄ Freeze: what is right no longer moves

A box has come out right: the product, the character, the title? Select it and press ❄ Freeze what was rendered here: what the render shows inside that box becomes a locked layer, on top. Change the seed, the scene, the other boxes, run again: the scene changes around it, the layer comes back identical to the pixel.

On an object box, you freeze the object (cut out); on a text box, the letters.
☀ Unfreeze makes the box generative again: it will be redone on the next render.

❄ Background (in the mask bar) freezes the whole image as background: the setting no longer moves, and your inferences happen on top of it. The button stays lit while the background is frozen; ☀ Background (or Unfreeze on the background) releases it.

> The frozen background takes the entire render, texts included: a text moved afterwards leaves its old letters in the background.

4 · The mask: redo one area, and only that one

Press ✎ Mask, then paint on the image: a single stroke opens a zone and starts painting it. Each zone has its own instruction (the Mask: field). Paint over a box: the zone belongs to it; a stroke started outside any zone opens a new one.

You infer one zone at a time: select the zone, then ⟳ Infer mask. The result is added to the displayed image; the next zone starts from there. You edit the image zone after zone, losing nothing from the previous one.

⟳ Infer box: redoes the whole selected box, without painting.
The gesture: brush → zone and ellipse redo the rectangle that surrounds what you painted; rectangle draws a rectangular zone, which you move and stretch like a box; free brush redoes only the painted shape, with its strength, and an adjustable soft edge (edge, in pixels, outwards, inwards or centred).
The intent: inpaint fills the zone with what you ask for · refine reworks what is there · remove erases and rebuilds the background.
The model:
model 1: place an object, a subject, an effect (fire on a title);
model 2: transform what is there, its material;
model 3: alter what the zone shows, without replacing it: bigger, darker, open; the dress stays a dress;
model 4: enhance the material (wrinkles, hair, textures), with no instruction.
⌫ Eraser (or Alt while painting) erases the mask · ✕ Clear mask removes the masks.

On a text, the letters keep their shape. A zone placed on a text (even if started next to it) puts the material inside the letters: fire letters, without ever redrawing them. Flames that spill over go into the background.

Deleting a zone whose result is finished keeps it in the image; a zone not yet inferred leaves without a trace.

[screenshot: FOREST in letters of fire, before / after]

5 · Placing the exact

A placed image is never redrawn. Drop it on the board, or ▣ Image then Choose an image...

✂ Remove the background: it becomes a transparent cut-out, with no rectangle around it (...keep every subject keeps all the subjects). ✂ Auto cut-out in the toolbar cuts out every image you drop.
⛶ Fill the picture: it covers the whole image, edge to edge, proportions kept; whatever overflows is cropped at render. ▣ Fit in the picture: it fits entirely, centred.
◇ Distort / turn: four corners to drag: turn it, stretch it, set it in perspective on a surface (a plate, a sign, a screen). Shift + corner resizes it within its perspective; ▭ Back to a rectangle cancels.
Effect: engraved (cut into the surface) or embossed (raised), with its depth; integrate redraws it in the light of the scene; integrate + paste back does both.

Exact text: a style (logo, title, bold, medium, body) or a typeface, alignment, tracking, leading, colour. The size is the largest that fits inside the box.

6 · Layouts: keeping a composition

A layout is your boxes: positions, descriptions, texts, styles, images. In Layout:

type a name and 💾 Save: it is stored in your layouts, and the Layouts... list reloads it on any image;
⤓ Export downloads it, ⤒ Import reloads it (Ctrl+Z undoes). A layout is yours: you export it, you give it away.

Build a poster once, place it on the next image.

On the board

↖ (V): dragging a box moves it (Shift locks an axis, Alt copies it) · 8 handles to resize it · X, Y, W, H to the pixel in the box settings.
Arrow keys: move (Shift ×10) · Tab: next box · Delete: delete · Esc: deselect.
Ctrl+Z / Ctrl+Y: undo / redo · Ctrl+D: duplicate · Ctrl+C / Ctrl+V: copy boxes.
Shift+click adds to the selection, Ctrl+A selects everything. Ctrl+G groups, Ctrl+Shift+G ungroups.
Drag from an empty spot (or Ctrl+drag from anywhere): rectangle selection.
An image that covers the whole frame is selected with one click; dragging over it remains a rectangle selection.
◐ Render shows / hides the render under the boxes · ▢ Boxes the boxes · ▦ Grid the grid · ⬚ Margin a margin around the image (whatever overflows is cropped at render).
The alignment bar: align, distribute, same width / height.

The box list: 🔒 lock · ◉ switch on / off (a box switched off stays on the board but is sent nowhere) · ▲ ▼ or drag the row for the order (top = in front) · ✕ delete.

The render

Click the render to open it full size. From there: upscale (Upscale), edit (Edit), enhance the material. Every render goes into your creations.