Project Eon: Multimodal Time-Machine

Inspiration

We were inspired by a simple question: What if the same city existed in a completely different era?
Modern architecture reflects technology, culture, and resources of today. We wanted to build an AI system that can reimagine a real location (like Toronto Waterfront) as if it belonged to a speculative civilization such as Industrial Rome or Solarpunk Maya. This project combines creativity with logic by using AI to reconstruct believable alternate histories.

What it does

Project Eon is an AI-powered architectural time machine.
It takes a modern photo of a city, building, or landmark and generates a reconstructed version in a selected speculative era.

It performs two major tasks:

Forensic audit of the image
Detects architectural elements like materials, building style, skyline shape, lighting, and layout.
Era-based reconstruction
Transforms the image into an alternate historical or futuristic style while preserving the scene structure.

Example outputs:

Toronto Waterfront → Industrial Rome
Downtown skyline → Solarpunk Maya
Modern street → Cyberpunk Victorian

How we built it

We built Project Eon as a multimodal AI pipeline:

Image Understanding
- Used computer vision to detect key structures (buildings, roads, water, sky).
- Extracted scene features such as geometry, edges, and dominant materials.
Era Prompt Engine We designed an "era logic" layer that maps modern features to era-specific replacements.
For example:
- Glass skyscrapers → stone towers / bronze structures
- LED lighting → torches / solar lanterns
- Concrete bridges → arches / wooden rope bridges
Image Generation
- Used diffusion-based image generation to produce a new version of the scene.
- Added constraints to preserve the original layout and perspective.
Final Output System
- Frontend: Web interface where users upload a photo and choose an era.
- Backend: API that runs image analysis + generation.
- Database: Stores uploaded images, eras, and generated results.

The process can be summarized as: [ Output = G(Image, Era, Constraints) ] where (G) is the generative model and constraints preserve structure.

Challenges we ran into

Maintaining realism: Generated images sometimes looked artistic but not logically believable.
Preserving geometry: Some transformations distorted the original perspective.
Era consistency: Keeping every object aligned with the chosen era was difficult.
Compute cost: Image generation is GPU-heavy and slow on normal systems.
Prompt engineering: Small changes in prompts caused big changes in output.

Accomplishments that we're proud of

Built a complete multimodal system (vision + generation).
Achieved consistent and high-quality reconstructions for multiple eras.
Created an interactive UI that feels like a real product.
Developed an "era logic" mapping system instead of only relying on random prompts.
Produced outputs that are both creative and historically-inspired.

What we learned

Multimodal AI is much more than image generation—it needs logic and constraints.
Good results require a balance between creativity and structure.
Prompt engineering is a real technical skill, not just trial-and-error.
Architecture has patterns that AI can learn and reinterpret.
Deploying AI is challenging because GPU resources matter as much as code.