Houzz — AutoMate AI: 3D Products
Unlocking Design Freedom for Pros
I designed an image-to-3D feature powered by Houzz's AutoMate AI, letting pros generate any 3D product needed to fully realize their clients' home renovation projects.
web
3D
ai
3D Models Generated
6,500+
Pros can't deliver their client's vision with our small catalog of 3D products
Our team receives 1,000+ monthly requests from pros asking for specific 3D furniture and appliances to be added to our catalog—exceeding what our two 3D artists can manually fulfill. As a last resort, pros are using generic placeholders in their floor plans, which don't convey a real design vision and reduce client confidence.
Example of a room completely furnished with generic products due to the lack of real furniture options
Opportunities
Following strong adoption of our recent 2D-to-3D floor plan AI conversion feature, we set out to extend our AI capabilities to generate products from images.
🎨
Unlock design freedom
Enable pros to furnish their floor plans with any product using an AI-powered solution that converts 2D images into 3D products.
🌱
Crowdsource our 3D catalog growth
Turn pros into contributors by allowing them to share their generated models with the Houzz Pro community—scaling our 3D catalog exponentially.
🚀
Stand out from competition
Houzz Pro would be the first to let pros generate 3D products without any specialized software or experience, saving them time and money.
Process
I collaborated with PMs and engineers to shape our MVP requirements, drawing inspiration from generative AI and 3D tools to conceptualize our vision.
Our research engineer's proof of concept showed promising results but also exposed many limitations with our AI. Technical constraints and cross-functional platform dependencies added additional complexities to our project that required us to design alternative solutions that balance tradeoffs between conversion quality, usability, and feasibility.
Working around our AI's strengths and limitations, I designed a guided, step-by-step experience with several guardrails that help guide both pros and the AI towards more reliable results. After multiple design reviews and iterations, we ran beta tests and refined the experience for launch.
A snapshot of my design proposals addressing technical limitations with multi-layered 3D canvases
Early AI conversion results produced low quality results, but significantly improved as we tuned our technology
Designing the AutoMate AI workflow
Our step-by-step approach balances automation with user control, building trust in both the tool and the visualizations they share with clients. AI limitations influenced the workflow, which we broke into three major steps: sourcing & image selection 🡢 image segmentation 🡢 product detail verification.
Overview of the image-to-3D workflow
Step 01
Sourcing & Image Selection
Our system needs a lot of product metadata to get started (e.g. name, type, dimensions). While we support manually uploading images and filling out these details, the process can be time-consuming and error prone.
To streamline this step, we integrated our Houzz Clipper tool—a browser extension that automatically exports product images and specifications into Houzz. This tool not only saves pros time, but ensures more accurate data is passed directly to the AI. By driving the adoption of our own Clipper tool, we're also reinforcing platform dependency.
With a single click, our Clipper tool saves product specs directly into Houzz Pro to use in projects
Step 02
Image Segmentation
Our proof of concept revealed that our AI struggled with images containing multiple objects—it couldn't identify what to convert, leading to low-quality results. Since many pros already had low trust in AI tools, it was important to deliver high-quality results from the start to build credibility and retention.
We introduced a "segmentation" step where pros isolate the product from its background, giving the AI a cleaner image to work with. Prioritizing feature adoption and engagement, we decided on a more manual approach for MVP, trading pure automation for precision and flexibility. This led to higher quality results and helped build customer trust in the AI.
During our initial beta tests, we found that customers misunderstood the segmentation task: many thought they needed to select what to remove, not what to include. To improve task success rate, we added animations and tooltips to educate pros on better segmentation techniques.
Some beta testers thought they were supposed to select and remove the background, which produced results opposite of what they wanted
Guiding pros through the segmentation task
Step 03
Product Detail Verification
Even with clean segmentations, the AI sometimes produce results that were misaligned or incorrectly scaled. Additionally, our system doesn't know how the model should behave—whether it sits on the floor or attaches to walls or ceilings.
To address this, we created a two-step review flow:
Product orientation: review or correct the model's default alignment
Product details: validate or enter attributes like product type, dimensions, and behavior
I optimized these steps around our Clipper tool, importing more reliable attributes. Because these attributes can't be edited after saving, I added extra visual elements in the 3D canvas to help pros confidently validate the results before publishing.
Product orientation and details need to be added or verified for accuracy before saving to our library
A 3D scene created with AI-generated products
Scaling our 3D product catalog
While the image-to-3D workflow addressed pros’ immediate needs, we saw a longer-term opportunity: turning pros into contributors.
Each generated model was added to our shared catalog, instantly benefiting the entire Houzz Pro community. This approach not only scaled our 3D product catalog with minimal effort from our team, but also improved our AI's performance by preventing duplicate conversions and reducing operational costs.
Pros can now discover and use any user-generated 3D model from our catalog
Launch & impact
After eight months, our newest AutoMate AI for 3D Products feature was launched in April 2025. Within two months, we have:
Enabled pros to generate 6,500+ models—helping them create more realistic floor plan visualizations and improve client relationships
Crowdsourced 1,800+ new products into our shared 3D catalog—rapidly scaling our product library with minimal internal effort
Delivered one of the first AI-powered product conversion tools for floor planners—differentiating Houzz from competition
"Pros want their 3D floor plans to include the exact products that they are specifying, and Houzz Pro now enables them to do so by turning a 2D product photo into a 3D model."
Alon Cohen, Ceo, Houzz
Reflection & next steps
3D Floor Plans is the most popular tool in Houzz Pro and I feel very lucky to head design efforts on such a high-impact product. Working on this project has given me a few takeaways:
🛑
Sometimes friction is good
Early on, I kept thinking "simplify, simplify, simplify" and pushed for a fully automated, hands-off experience that everyone expects from AI features. After realizing that pros weren't reading instructions or verifying their results, I reintroduced friction to slow them down—leading to higher quality conversions.
🤖
AI can't replace humans (yet)
In an industry where attention to details really matter—especially for interior designers, it's still important to have a human touch. This reinforced our belief that AI is a tool that enhances a workflow.
As for next steps, pros are pushing the boundaries of our feature, such as uploading materials, flooring, landscaping, and even people. We're actively exploring ways to support some of these use cases, while also setting clearer guidelines. I'm excited to continue expanding our AI capabilities and help pros bring their home design projects to life.