Deep Generative Models with Hard Linear Equality Constraints
Ruoyan Li, Dipti Ranjan Sahu, Guy Van den Broeck, Zhe Zeng · 2025
Read Deep Generative Models with Hard Linear Equality Constraints by Ruoyan Li, Dipti Ranjan Sahu, Guy Van den Broeck, Zhe Zeng (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-05416.txt.
Description
Ruoyan Li and co-authors wrote a paper from 2025 on teaching image-making models to obey hard rules. Such models learn the look of a data set and then dream up new samples. They tend to break firm rules that encode real-world knowledge, such as a law that parts must sum to a set total. Past fixes used rough tricks and often warped the output away from the data's true look.
The team's method is built on sound chance math. It learns the data spread as seen through the lens of the rules, so every sample it makes obeys them by design. New gradient gauges make that lens shape trainable end to end. They tested the method on five image sets and three science tasks with straight-line rules. Plain models broke the rules near always. Their method kept every rule and beat rival fixes on realism.
Catalog record
| CREATOR | Ruoyan Li, Dipti Ranjan Sahu, Guy Van den Broeck, Zhe Zeng |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 80 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-05416