HorizonWind2004/reconstruction-alignment

[ICLR 2026] Official repo of paper "Reconstruction Alignment Improves Unified Multimodal Models". Unlocking the Massive Zero-shot Potential in Unified Multimodal Models through Self-supervised Learning.

45
/ 100
Emerging

This project helps AI model developers enhance the performance of their unified multimodal models (those that handle both text and images) by applying a technique called Reconstruction Alignment (RecA). By integrating RecA during the self-supervised training phase, developers can improve their models' zero-shot capabilities across various tasks like image generation and editing. The input is an existing multimodal model, and the output is a significantly improved version of that model, ready for deployment.

378 stars.

Use this if you are developing or deploying unified multimodal AI models and want to boost their performance and capabilities without extensive new training data or complex reinforcement learning.

Not ideal if you are a general user looking for an off-the-shelf application, or if you are not working with multimodal AI model development.

AI model development multimodal AI image generation image editing zero-shot learning
No Package No Dependents
Maintenance 10 / 25
Adoption 10 / 25
Maturity 15 / 25
Community 10 / 25

How are scores calculated?

Stars

378

Forks

15

Language

Python

License

Apache-2.0

Last pushed

Mar 13, 2026

Commits (30d)

0

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/diffusion/HorizonWind2004/reconstruction-alignment"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.