⏳ Awaiting GPU hardware
🏠 Floor Plan Vision AI
Read an architectural floor plan and describe its rooms, layout, circulation and dimensions.
The Gradio demo is written and committed to this repo — app.py,
requirements.txt and five example plans are all here. It needs a
GPU attached before it can run, so the Space is published as a static page
in the meantime.
What it does
Upload a floor plan, ask a question, and the model returns a written analysis: room count, how spaces connect, approximate dimensions, and notable architectural features. It reads standard plan symbols and dimension annotations.
Example plans included
The model
- Adapter — sabaridsnfuji/FloorPlanVisionAIAdaptor, a LoRA (r=16, α=16) over both the vision tower and the language model.
- Base — Llama 3.2 11B Vision Instruct, quantized on load to NF4 with double quantization and bfloat16 compute — matching the recipe the adapter was trained against.
- Serving —
transformers+peft, streaming token-by-token, targeting ZeroGPU.
Turning it on
Once GPU hardware is attached, swap the README frontmatter to:
sdk: gradio
sdk_version: 6.26.0
app_file: app.py
python_version: "3.12"
startup_duration_timeout: 1hNo code changes are needed.