How to write prompts for AI retouching

AI retouching edits a finished visualisation based on a text description, the prompt. Three things affect the result: the source visualisation, the model and the prompt text. This article collects tips to help you get the edit you need on the first or second attempt.

Source visualisation

The neural network works with the image you give it, so the quality of the source directly affects the result.

• Run retouching on the original visualisation. When an already edited image is processed again, details get lost and errors build up.
• The higher the resolution and detail of the visualisation, the more accurate the result.
• It is convenient to find the right wording at 1K and produce the final image at 4K. The result will differ slightly, but finding a working prompt costs less.

Choosing a model

Select the model from the drop-down list below the prompt field. T wo models are available:

• Ando follows instructions precisely. Choose it when you need to replace a specific item, change a material or add people while keeping everything else intact.
• Gaudí interprets the prompt more freely and costs less. It suits exploring options and mood, when changes elsewhere in the frame are acceptable.

The cost in tokens depends on the model and resolution. It is shown on the Generate button before you start.

Prompt presets

A preset contains a prompt we have tested on real visualisations. The first line of a preset starts with // and tells you what you can change in it.

• Edit the first paragraph: what to replace, what with, and where it is in the frame.
• The remaining paragraphs protect the geometry, lighting and scene. Avoid deleting them or adding protections of your own: each addition usually makes the result worse.

Add furniture from photo

Attach a photo of the item using Add images. Name what to replace or where to place the new item, for example "the armchair by the window". Set the size by comparison: "the same width as the current armchair".

Improve textures

Name the item and its position, then the material and colour, for example "the curtain to the left of the window, light linen". You can attach a photo of a fabric sample.

Add silhouette

State how many people are in the frame, where they are and what they are doing, for example "two people sitting at the dining table".

How to write your own prompt

• Name the item together with its position. Write "the floor lamp to the right of the sofa". Without this anchor, the neural network will put the item in the nearest free space.
• Make one edit at a time. Replace the view through the window and change the room lighting in two separate generations.
• Set the size using nearby items. Models ignore centimetres but handle comparisons with items in the frame well.
• Describe what you want to get. Prohibitions backfire: after "don't change the shadows", the model most often changes exactly the shadows.
• Mention lighting only for the new item, for example "with a soft shadow on the floor". General phrases about scene lighting redraw the whole image.
• Keep it short. Three to five sentences hold the scene better than a long text.
• Use original brand and model names: Verpan VP Globe, Eames Lounge Chair. The neural network recognises them more accurately than descriptions.
• Attach one image. One image is more reliable than two. In the prompt, say what to take from it: "replace the armchair by the window with the armchair from the attached photo" .

Example. Instead of "Change the sofa to a nice one", write: "Replace the grey sofa by the window with a light bouclé sofa of the same width. Give it a soft shadow on the floor."

If the result is not right

• Too much has changed. Shorten the prompt and remove prohibitions. Try the Ando model.
• The edit was not applied. Specify where the item is. Check that the prompt contains only one edit.
• The item is the wrong size. Compare it with a nearby item in the frame.
• The lighting changed across the whole room. Remove general words about lighting from the prompt.
• Hands or small objects are distorted. This is the hardest area for neural networks. Run one or two more generations.