AI transparency

VEDUTUM uses AI models in three places: reading a plan, making an image and making a clip. This page says what they may do, what they may not, and what happens to your files. Last updated: 05 November 2026.

Two routes to an image

Route 1, renderer: you upload a finished model (IFC, glTF, GLB). A physically based renderer computes stills and camera flights from it. Nothing is added here: what is not in the model is not in the image.

Route 2, plan: you upload a floor plan as PDF or image. A language model reads walls, openings, rooms and scale from it; VEDUTUM builds a model you check and correct in the browser. Only that model, rendered as a plain view, goes to an image model together with your choice of facade, frames and roof.

What the image model may and may not do

• Fixed: camera position, outline, storeys, roof shape and the position and size of every window and door. These come from your plan, not from the model.
• Added: material and surface of the chosen facade, light and sky for the chosen time of day, planting, paving and indicated neighbouring buildings.
• Therefore: an image produced this way is a design representation, not a construction document and not a promise about material, planting or surroundings. Check that it matches the permitted state before publishing.
• Labelling (Art. 50 EU AI Act): every image and video from the image or video model leaves our server marked in a machine-readable way: XMP metadata to the IPTC standard (DigitalSourceType = trainedAlgorithmicMedia) and EXIF fields, videos in the container metadata. In the project and on the share page the line “AI-generated visualisation” stands under each such image and video. Panoramas, sun studies and model exports are computed from the geometry without an AI model and are called model views. Whoever publishes the pictures, in a brochure or on a website, is obliged under Art. 50 (4) to label them as AI-generated where they could be taken for a photograph of a real place; the metadata stays in the file for that. If you want the mark “KI-Visualisierung” burned into the corner as well, tell us and we switch it on for the account. A C2PA signature (Content Credentials) is prepared and follows as soon as we hold a certificate for it.

Which models we use

• Recognition, images and clips: specialised models from several providers, through paid interfaces. Which ones exactly we do not publish — it is part of our work and it changes as soon as a better model appears. What the models may and may not do is set out above and holds whichever one is computing.
• Website text: drafts may be written with a language model. Every text is reviewed and signed off by a person before publication; no draft is published automatically.

For processing, the plan page, the model view and your free-text notes for the image model leave our servers and go to those interfaces. We only use paid access, for which the provider excludes the use of submitted content for training its models; we do not use free tiers, where content is evaluated for product improvement. Processing is not bound to a region; the provider stores inputs for a limited time to check for abuse and does not use them for training. We do not pass on account data. Whatever is printed on the uploaded plan page does go along, though — a title block with the client, the address, the planner or a stamp, as well as your notes for the image model. Crop or black out the title block before uploading if it does not belong there. We do not store the view sent to the model. The instruction and settings of every call stay in the project as a log for twelve months, so that points and results remain traceable. Anyone who needs the processors by name — for a record of processing activities, say — finds them in the privacy notice (German), which is where they belong.

On this website

• Images and clips: all of them show the same in-house study, “Haus Lindenweg”, produced from a floor plan we drew ourselves, using exactly the sequence described above. No built house, no client project, no stock photography.
• Demo: the demo on the home page computes its 3D model in your browser. Images and clips inside it are the finished results of that study; the demo calls no AI model and sends nothing to us.
• Chat: there is no AI chat on this website.

Contact

Questions about our use of AI: office@vedutum.com