Before GPT Image. Before Nano Banana.
Starting in March 2024, we built MaiProperty around segmentation, Stable Diffusion inpainting, ComfyUI integration and AWS GPU infrastructure. The product connected lawn designs to marketplace products, quotations, contractor leads and lender pre-approval: a turnkey makeover journey with a custom AI pipeline underneath.
The timing matters
We built the workflow
before those models arrived.
In March 2024, our team was choosing segmentation models, tuning masks and comparing inpainting checkpoints. Preserving a house while redesigning its lawn required explicit control over the image-processing stages. ComfyUI, a Python service and AWS GPU compute brought that work into the application.
GPT Image 1 reached the API in April 2025. Nano Banana followed in August 2025. MaiProperty's development began the year before both.
March 2024
MaiProperty research begins
Segmentation + inpainting
23 April 2025
GPT Image 1 API26 August 2025
Nano Banana
Change the lawn. Keep the property.
A homeowner needed to see an alternative for their own yard, with the surrounding building and boundaries still recognisable. Our research treated that as two linked problems: first isolate the editable lawn, then condition the inpainting model on the image, mask and requested design.
A useful picture was only the beginning. Customers also needed suitable products, a quotation, a service provider and a way to explore financing. We built the customer experience and the business workflows around that complete journey.
All-object segmentation produced many regions. The pipeline needed the lawn, with control over mask boundaries and surrounding trees.
Early SDXL tests changed details outside the mask. Checkpoint selection, guidance, denoising and prompt settings all mattered.
The design had to connect to marketplace sourcing, quotations, contractor leads and lender connections.
The pipeline was part of the product.
We worked across the image-processing workflow, the application backend and the service layer. Each stage carried the customer's property and preferences forward.
GroundedSAM targeted the lawn using a text prompt, with trees excluded and mask boundaries adjusted before inpainting.
We evaluated Stable Diffusion and SDXL inpainting checkpoints, prompts and inference settings. A Python API connected the ComfyUI service to the application; the backend later integrated an SD3 endpoint.
Third-party marketplace product matching connected the visual direction to materials and sourcing options. Products and proposed landscaping work fed the quotation journey.
Vendor discovery used service and city filters. Customers ready to proceed could become leads for relevant vendors and contractors.
Lender connections supported loan pre-approval for lawn makeovers, helping eligible customers explore how to fund the proposed work.
Customer histories, collections and feedback sat alongside vendor portfolios, subscription billing and admin controls for products, services and AI prompts.
The product experience
The design was the start.
The service journey came with it.
A project with context
The Express backend assembled service-specific prompts, called the inference API, saved the generation and returned relevant vendors and products.
The business behind the image
Vendor portfolios, billing and subscriptions worked alongside admin controls for customers, products, services and AI prompts.
A visual idea with a path to delivery.
The result was an integrated product and service journey. A customer's photo could become a saved design, a product selection, a quotation and a conversation with the businesses needed to carry out the work.
From model experiments to an operating platform.
Segmentation and inpainting research
Compared segmentation approaches, tested lawn masks, and evaluated inpainting quality, inference settings and GPU processing time.
Application and business portals
Built the customer generation flow, Express orchestration, Supabase persistence, vendor and admin workspaces, and subscription workflows.
Inference integration evolved
Updated the application's generation API integration to an SD3 endpoint. The inference service remained separate from the web application.
Deployment and continued support
Supported AWS environment migration, inference-domain and SSL configuration, and technical handover for the ComfyUI service.