Generative image tools stop at the picture. Flat Dream trains its own models on a specific architectural language, then keeps going, into a magazine builder, a mesh converter and a model viewer, so the output stays usable.
Context
IAAC · Barcelona
Year
2026
Stack
LoRA training · ComfyUI · LM Studio · Interface design
Brief
FlatDream is an interface inspired by the book Learning from Las Vegas, built on 2 custom-trained LoRA models, fine-tuned on the visual language of the Archigram movement. It connects directly to ComfyUI and LM Studio to generate architectural imagery from text, image references, and multi-input compositions, with outputs that feed into a magazine builder, 3D mesh converter, and model viewer.
01 · Context
A general image model has no particular architecture in it. Flat Dream starts from a stated position (Archigram's visual language, read through Learning from Las Vegas) and fine-tunes two LoRA models on it, so the output argues from somewhere.
02 · System
The interface drives ComfyUI and LM Studio directly, generating from text, image references and multi-input compositions. Running against local models keeps the loop fast enough to iterate in, and keeps the work off someone else's server.
03 · Outputs
Generated imagery feeds a magazine builder, a 3D mesh converter and a model viewer. That last step is the point: an image is a mood, a mesh is something you can measure, and the tool carries the idea across that line.
