Image Generator

Imagine being able to generate an image from pure math. No libraries.

Comment

Latent Flow Matching Image Generator

Live Demo: Check it out here

Source Code: Github

What it does

We built a generative image model from scratch using Latent Flow Matching. By combining a Variational Autoencoder (VAE) with a lightweight MLP velocity predictor and vanilla classifier-free guidance, our model can effectively sample and generate clean images that capture the underlying MNIST data distribution.

How we built it

  • Model Architecture & Training: We built both the VAE and the MLP velocity network completely from scratch, writing a custom training loop to learn the MNIST data distribution.
  • Compute: We leveraged Hyak Klone (high-performance computing resources) to train the model and extract our final model weights.
  • Full-Stack Interface: We developed a web interface backed by a Flask server that bridges the frontend (HTML/CSS/JS) with our backend model, allowing users to interactively prompt and generate images seamlessly.

Challenges we ran into

  • Compute & Time Constraints: Training deep generative models from scratch always eats up precious hackathon hours, leaving little room for hyperparameter tuning.
  • Integration Overhead: Wiring up the Flask backend with asynchronous JavaScript and HTML front-end communication took a significant chunk of time away from core model experimentation.

Accomplishments that we're proud of

Successfully training a custom flow-matching model from scratch on cluster compute and wiring the entire pipeline, from raw GPU weights to a live, interactive web demo.

What we learned

We gained a deep, hands-on understanding of how to set up robust Flask servers and manage stateful, asynchronous communication loops between Flask <-> JavaScript <-> HTML.

What's next for Latent Flow Matching

  • Scaling Up: Transitioning from our lightweight MLP to a Multimodal Diffusion Transformer (MMDiT) architecture to handle general, high-resolution imagery.
  • Cloud Infrastructure: Shifting compute entirely to scalable cloud instances rather than local web execution to handle larger parameter sizes and deliver a lightning-fast, state-of-the-art user experience.
Share this project:

Updates

Submission history