Skip to content
Other

hackathon-team6-cubicasa-ae

reverse image search for similar floor plans

2 min readupdated

A convolutional autoencoder (an AI model that compresses an image into a compact numerical vector and reconstructs it back) built by a student team from IAAC MaAI 2026/27. The model encodes a 128×128 pixel architectural floor plan into a 128-number code: layouts with similar shapes and proportions end up close together in vector space, enabling instant similarity search. The repository already includes precomputed vectors for 5,000 real house plans from the CubiCasa5K dataset, while the model itself weighs just 4.5 MB in ONNX format (an open standard for fast AI model runtime) and runs smoothly on a standard CPU. It is a handy tool for architects and students looking to quickly source references based on overall building geometry.

Key features

  • Ultra-lightweight architecture: 1.17M parameters with ONNX and PyTorch weights, allowing training and inference without a heavy GPU.
  • Dual outputs: returns both a 128-dimensional code vector for distance-based search and a reconstructed reconstruction image.
  • Ready-to-use search database: bundles precomputed collection_codes.npy embeddings covering 5,000 mostly Finnish residential floor plans.
  • Outperforms classical baselines: when a plan is shifted by 3 pixels with added noise, it retrieves the original 69% of the time, compared to 25% for PCA and 4% for raw pixel matching.

What you can do

  • Find architectural analogs: upload a sketch or blueprint to pull the top 5 CubiCasa5K floor plans with matching contours and proportions.
  • Visualize and cluster collections: generate UMAP maps of large blueprint archives, grouping layouts by aspect ratio, footprint area, and line density.
  • Embed into web services: drop the 4.5 MB ONNX file into a local app or Telegram bot for instant floor plan reference lookup.

Example

session = ort.InferenceSession(hf_hub_download(repo, "floorplan_ae.onnx"))
collection = np.load(hf_hub_download(repo, "collection_codes.npy"))
ids = json.load(open(hf_hub_download(repo, "collection_ids.json")))

code, reconstruction = session.run(None, {"image": preprocess(plan_path)})
distances = np.linalg.norm(collection - code, axis=1)
for i in np.argsort(distances)[:5]:
    print(ids[i], round(float(distances[i]), 1))
Read moreDetails
Metrics (CubiCasa5K test split, 400 plans): Test MSE 0.0131 (vs. 0.0153 for the 32-dimensional version), Retrieval top-1 at a 3 px shift — 69% (up to 98% on high-contrast scans, 28% on faint ones).
Strict preprocessing pipeline: grayscale conversion, color inversion (white walls on a black background), black padding to a square, Lanczos resize to 128×128, and [0, 1] normalization.
Limitations: 128×128 resolution drops doors, furniture, and text labels; the model is not rotation-invariant and remains sensitive to scan line brightness.
License: CC BY-NC-SA 4.0 (non-commercial use only).
Examples
Discuss Telegram 3

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.0 GB · ready files0.5 GB
  • FP8 / INT8recommendedweights 0.0 GB0.5 GB
  • 4-bitweights 0.0 GB0.5 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # iaac-maai/hackathon-team6-cubicasa-ae
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download iaac-maai/hackathon-team6-cubicasa-ae --local-dir ~/models/hackathon-team6-cubicasa-ae
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of hackathon-team6-cubicasa-ae (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

iaac-maai/hackathon-team6-cubicasa-ae
Author
iaac-maai
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

03

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments