This working prototype runs two models in sequence. A short written prompt becomes a single concept image of a house on its land, and that image is then converted into a concept-stage 3D mesh that can be orbited and zoomed. The test covers the complete path from words to image to form.
Two models do the work, each kept to one job. The first is Google's Gemini image model, which turns the prompt into a clean, centred concept render on a plain background. The second is TRELLIS, an image-to-3D model, which lifts that render into a mesh loaded straight into the viewer. The prompt carries the intent, the image model carries the look, and the 3D model carries the geometry.
This is concept-stage only. The result is a visual mesh without measured walls, rooms or dimensions, and the ground is generated rather than taken from a real site. It can help a person assess the general form and relationship to the landscape. Accurate design work still requires a measured model and professional development.
A quick way to see a form, not a measured model. The further you push it toward documentation, the more a professional has to own.
Generation runs through the site, calling fal.ai for both the image and the 3D step, so it needs a key configured in the server environment. To keep the demo usable for everyone, each visitor can run it a limited number of times per day. With no key set, the tool still loads, and the generate steps return a clear message instead of a result.