Implicit Maximum Likelihood Estimation

Ke Li, Jitendra Malik

Link to Paper
Reviews
Slides

Other Related Papers:

Super-Resolution via Conditional IMLE
Diverse Image Synthesis from Semantic Layouts via Conditional IMLE
On the Implicit Assumptions of GANs

Slides

The slides (which includes an illustration of how the algorithm works) can be found here.

Code

The code can be downloaded here.

Applications

Our method can be used to train implicit probabilistic models (a common example being the generator in GANs). Unlike GANs, however, our method does not suffer from mode collapse/dropping and is stable to train. As a result, we are able to generate different predictions for the same input. Below are two applications that take advantage of this capability.

Super-Resolution

Below are the different high-resolution images generated by our method for the same low-resolution image:
Input:
Input
Output:
Output

Image Synthesis from Semantic Layouts

Below are the different images generated by our method for the same scene layout:
Input:
Input
Output:
Output