GenPlan : Generation Vector Residential Plans Based on the Textual Description [RU/EN]

,

Contacts: Egor Bazhenov

Scheme of the method

We present an algorithm for generating 2D/3D vector residential plans based on a text description. This algorithm supports several output data options. SVG is supported for 2D. IFC, OBJ, glTF, and GIF are supported for 3D.

Generation example

Text prompt	Raster Image	Vector Image	3D Image
A minimalist 2D floor plan of an empty studio apartment without any text, featuring clean black lines and white spaces, showcasing an open layout with designated areas for living, sleeping, and kitchen. without words on a picture

Usage

1. git clone https://github.com/CTLab-ITMO/GenPlan
2. pip install requirements.txt
3. python pipeline.py --text <your text promt> --generation_model <model_name> --result_type <2d or 3d> --formats <svg, gif, gltf, ifc, obj> --control_image <list of coordinates pairs>

Script launch parameters

Parameter name	Description
--text	Text description of the plan.
--generation_model	Model for generation. Can be one of these. SDXL is used by default.
--result_type	The expected type of result. It can be 2D or 3D. 2D is used by default.
--formats	Expected result formats. For 2D, only SVG. For 3D - all formats. SVG is used by default.
--control_image	List of 2d contour coordinates of the generated plan. None is used by default.

Setting parameters

Individual settings can be configured by config file for your task.

Parameter name	Description	Type
PNG_PATH	Generated bitmap image path	String
CLEAN_PNG_PATH	Cleaned bitmap image path	String
BLACK_COLOR_BORDER	Color filtering maximum value	Float
MAX_PERCENTILE	Maximum pixel differences percentage	Float
MAX_VALUE	Maximum value of dissimilar pixels	Int
MIN_THICKNESS	Minimum thickness of a vector line	Int
MAX_DEVIATION	Maximum deviation of points from a straight line	Int
LOSS_SCALE	Scale for a white loss	Float
EDGING_TYPE	Mask in which content would be, other space - white	String
USE_WHITE_LOSS	Whether to use white loss or not (works only with SDXL model type)	Bool

Reference

@conference{visapp26,
    author={Egor Bazhenov and Stepan Kasai and Viacheslav Shalamov and Valeria Efimova},
    title={GenPlan: Generation Vector Residential Plans Based on the Textual Description},
    booktitle={Proceedings of the 21st International Conference on Computer Vision Theory and Applications - Volume 3: VISAPP},
    year={2026},
    pages={283-290},
    publisher={SciTePress},
    organization={INSTICC},
    doi={10.5220/0014238300004084},
    isbn={978-989-758-804-4},
    issn={2184-4321},
}