Flanora AI/Flanora-AI-v1

Flanora-AI-v1 is a fine-tuned Canny ControlNet designed for controlled floor-plan image generation. It uses bedroom-count information extracted from a text prompt to select an appropriate Canny edge control image, which is then provided to a fine-tuned ControlNet together with Stable Diffusion v1.5 to generate a floor-plan image.

Model Details

Model Description

Developed by: Rana Talukdar (Rana-15), Ansh Gaur, Bindupautra Jyotibrat (BJyotibrat)
Shared by: BJyotibrat
Model type: Fine-tuned Canny ControlNet for floor-plan image generation
Language(s) (NLP): English
License: GPL-3.0
Finetuned from model: lllyasviel/sd-controlnet-canny, based on Stable Diffusion v1.5

Model Sources

Repository: https://github.com/Jyotibrat/Flanora-AI
Demo: https://flanora-ai.vercel.app/

Uses

Direct Use

Flanora AI can be used to generate floor-plan images from text prompts specifying a desired bedroom count between 0 and 5 bedrooms.

The inference pipeline extracts the requested bedroom count from the prompt, selects a corresponding Canny control image, and uses the selected image with the fine-tuned ControlNet and Stable Diffusion v1.5 to generate the final floor-plan image.

Example generated image:

Inference example:

Downstream Use

Flanora AI can be integrated into applications for:

Floor-plan image generation
Generative architectural visualization
Research into controlled image generation
Research and experimentation with ControlNet-based floor-plan generation

Out-of-Scope Use

Flanora AI should not be treated as a replacement for professional architectural design, engineering analysis, building-code verification, or construction documentation.

The generated images should not be interpreted as professionally validated architectural plans.

Bias, Risks, and Limitations

The model has several known limitations:

The training dataset is imbalanced across bedroom categories.
The dataset contains substantially more 3-, 4-, and 5-bedroom images than 0-, 1-, and 2-bedroom images.
The model does not reliably generate the requested bedroom count.
Performance is particularly weak for 4- and 5-bedroom prompts.
Generated floor plans may contain disconnected rooms, invalid walls, or problematic doors.
The model can generate visually plausible floor-plan-like images that are not necessarily architecturally valid.
Automatic image metrics do not fully measure architectural correctness.

In the held-out evaluation set, the model achieved a bedroom-count accuracy of 11.94% and an overall structural validity rate of 35.82%.

Recommendations

Users should treat generated floor plans as experimental outputs rather than construction-ready architectural documents.

Human or domain-expert review is recommended before using generated layouts for any practical architectural purpose.

How to Get Started with the Model

Use the provided inference.py and bedroom_selector.py files to run inference.

The inference pipeline accepts a text prompt containing a bedroom count, such as:

Generate a 3 bedroom floor plan

The bedroom count is extracted from the prompt and used to select the corresponding Canny control image before generation.

Training Details

Training Data

Flanora AI was trained using a custom merged floor-plan dataset combining:

ROBIN
CVC-FP / ImagesGT

The merged dataset is publicly available here: https://huggingface.co/datasets/BJyotibrat/ROBIN-ImagesGT-Merged-Flanora-AI-v1

The combined dataset contained 622 images.

Dataset distribution:

Bedroom Category	Images
0 bedroom	11
1 bedroom	19
2 bedroom	15
3 bedroom	221
4 bedroom	173
5 bedroom	183
Total	622

The original datasets remain attributable to their respective authors and are subject to their respective licensing terms.

Dataset Credits

ROBIN

D. Sharma, N. Gupta, C. Chattopadhyay and S. Mehta, "DANIEL: A Deep Architecture for Automatic Analysis and Retrieval of Building Floor Plans," 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), 2017, pp. 420-425, doi: 10.1109/ICDAR.2017.76.

@INPROCEEDINGS{8270007,
author={Sharma, Divya and Gupta, Nitin and Chattopadhyay, Chiranjoy and Mehta, Sameep},
booktitle={2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR)},
title={DANIEL: A Deep Architecture for Automatic Analysis and Retrieval of Building Floor Plans},
year={2017},
volume={01},
number={},
pages={420-425},
doi={10.1109/ICDAR.2017.76}}

ImagesGT

de las Heras, Lluís-Pere and Terrades, Oriol Ramos and Robles, Sergi and Sánchez, Gemma, "CVC-FP and SGT: a new database for structural floor plan analysis and its groundtruthing tool," International Journal on Document Analysis and Recognition, 2015.

@article{Heras15a,
author={de las Heras, Lluís-Pere and Terrades, Oriol Ramos and Robles, Sergi and Sánchez, Gemma},
title={CVC-FP and SGT: a new database for structural floor plan analysis and its groundtruthing tool},
journal={International Journal on Document Analysis and Recognition},
year={2015}}

Training Procedure

The floor-plan images were preprocessed using grayscale conversion followed by Canny edge detection.

The resulting edge images were used as ControlNet conditioning images.

Preprocessing

Input images were converted to grayscale.
Canny edge detection was applied.
Canny thresholds: 100 and 200
Images were resized to 512 × 512 during the training pipeline.
The dataset was organized according to bedroom-count categories.

Training Hyperparameters

Base model: stable-diffusion-v1-5/stable-diffusion-v1-5
Initial ControlNet: lllyasviel/sd-controlnet-canny
Optimizer: AdamW
Learning rate: 1e-5
Number of epochs: 3
Training resolution: 512 × 512
Batch size: 1
Gradient accumulation: 4
Mixed precision: fp16

Speeds, Sizes, Times

Training time: Approximately 0.5–1 hour
Evaluation and related processing: Approximately 1 additional hour

Evaluation

Testing Data, Factors & Metrics

Testing Data

The model was evaluated using a held-out test set containing 67 samples.

The evaluation included samples from the supported bedroom-count categories.

Find the full evaluation here: https://github.com/Jyotibrat/Flanora-AI

Factors

The evaluation was disaggregated by requested bedroom count:

0 bedrooms
1 bedroom
2 bedrooms
3 bedrooms
4 bedrooms
5 bedrooms

Metrics

The evaluation used the following metrics:

Edge IoU
Chamfer Distance
Edge SSIM
CLIP Similarity
Bedroom-count Accuracy
Room Connectivity
Wall Validity
Door Validity
Structural Validity
Multi-seed Consistency
Prompt Robustness

Results

Check all the evaluations here: https://github.com/Jyotibrat/Flanora-AI

Evaluation metrics:

Summary

Metric	Result
Mean Edge IoU	0.1341
Median Edge IoU	0.1316
Mean Chamfer Distance	4.3802
Median Chamfer Distance	3.9137
Mean Edge SSIM	0.7433
Mean CLIP Similarity	0.3071
Mean Consistency Edge IoU (Reliability)	0.1531
Prompt Robustness (Robustness)	0.3509
Bedroom-count Accuracy	11.94%
Room Connectivity	62.69%
Wall Validity	65.67%
Door Validity	73.13%
Structural Validity	35.82%

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

Hardware Type: NVIDIA T4
Hours used: Approximately 0.5–1 hour for model training; approximately 1 additional hour for evaluation and related processing
Cloud Provider: Google Colab

Technical Specifications

Model Architecture and Objective

Flanora AI is based on a Canny-conditioned ControlNet architecture.

The overall generation pipeline is:

The model was initialized from lllyasviel/sd-controlnet-canny, which is a Canny-conditioned ControlNet for Stable Diffusion.

The Stable Diffusion v1.5 base model used by the system is: stable-diffusion-v1-5/stable-diffusion-v1-5

Compute Infrastructure

The model was trained using Google Colab.

Hardware

NVIDIA T4 GPU

Software

The training and inference pipeline uses:

PyTorch
Diffusers
Transformers
Accelerate
Safetensors
OpenCV
Pillow
NumPy

Glossary

ControlNet: A neural network architecture that adds spatial conditioning to diffusion models.
Canny Edge: An edge-detection representation used as the structural conditioning input.
Edge IoU: Measures overlap between generated and reference edge structures.
Chamfer Distance: Measures the distance between generated and reference edge structures.
Edge SSIM: Measures structural similarity between edge representations.
CLIP Similarity: Measures semantic similarity between the generated image and its associated text prompt.
Bedroom-count Accuracy: Percentage of generated floor plans containing the requested number of bedrooms.
Room Connectivity: Percentage of generated floor plans in which all major rooms are reachable through logical paths.
Wall Validity: Percentage of generated floor plans with coherent major wall boundaries.
Door Validity: Percentage of generated floor plans with reasonably valid door/opening placement.
Structural Validity: Percentage of generated floor plans satisfying the defined connectivity, wall, and door validity criteria.