SnapSpace

Real rooms, real physics, unreal designs.

Inspiration

We've all tried AI interior design tools that look amazing but are completely impractical—turning a window into a painting or deleting a door entirely. We wanted to build a tool that respects the architecture of a home while reimagining its potential. The goal was distinct: Real rooms, real physics, unreal designs. We didn't just want a pretty picture; we wanted a renovation plan.

What it does

SnapSpace is an intelligent interior design agent that "sees" your room before redesigning it. Unlike standard filters, it understands geometry.

Analyzes Structure: Uses Google Gemini to detect immutable elements (walls, windows, beams).
Preserves Reality: Locks these elements in place using strict negative constraints.
Renovates Instantly: Generates high-fidelity, photorealistic redesigns in styles ranging from Minimalist to Cyberpunk, all without moving a single structural wall.

How we built it

We engineered a Hybrid AI Architecture to get the best of both worlds:

The Architect (Google Gemini 2.5 Flash): We used Gemini's vision capabilities not just to "see" the image, but to extract a logical "Geometric Manifest"—a strict list of constraints (e.g., "3 sash windows on left wall").
The Artist (FLUX.1 via Hugging Face): We piped these strict logic constraints into FLUX.1, a state-of-the-art image generation model, to simply "paint" the new aesthetic over the rigid framework provided by Gemini.
The Backbone (Python & Streamlit): Built as a lightweight web app that handles the complex orchestration between APIs seamlessly.

Challenges we ran into

The "Hallucination" Problem: Generative AI loves to be creative, often causing it to hallucinate new doors or move windows. We spent hours refining "Negative Constraint Prompts"—teaching Gemini to explicitly forbid the image generator from altering specific coordinates.
API Choreography: Managing authentication and connection states between Google, Hugging Face, and Pollinations was tricky. We had to build a robust "Fallback System" that automatically switches providers if one service goes down, ensuring the app never crashes.
Balancing Quality vs. Speed: We initially tried running heavy models locally (Flux.1-dev) but realized it excluded users without gaming PCs. We pivoted to a cloud-native approach to make SnapSpace accessible to anyone with a browser.

Accomplishments that we're proud of

Structural Fidelity: Achieving a workflow where the AI actually listens to "Don't move the window."
Smart Fallback System: The app is incredibly resilient; if the high-end Gemini model fails, it gracefully degrades to a faster, lighter model without the user even noticing.
Zero-Shot Analysis: It works on any room photo without needing training data or manual masking.

What we learned

Prompt Engineering is Programming: We learned that natural language prompts for AI need the same strictness as code. "Make it nice" fails; "Keep exactly 3 windows" works.
Hybrid AI is the Future: Combining a reasoning model (Gemini) with a generative model (FLUX) yielded far better results than asking a single model to do both tasks.

What's next for SnapSpace

3D Mesh Generation: Converting the 2D redesign into a basic 3D model you can walk through.
Shoppable Furniture: Identifying the AI-generated chairs and lamps and finding real-world purchase links.
AR Mode: Allowing users to hold up their phone and see the renovation overlayed in real-time.