EECS 545 Machine Learning 2021–2022
ICA-SeFa: Interpretable Image Controls
Course project with subsequent independent research
Archived course report (PDF)ICA-SeFa explores how to find understandable editing controls in a pretrained image generator. Building on SeFa, it applies independent component analysis to generator weights to discover directions that change properties such as expression, hair color, and lighting, without retraining the generator.
I proposed the method and its mathematical rationale, shared the implementation, and developed a CLIP-based evaluation of intended changes and effects on other attributes. The work began in EECS 545 in 2021 and continued independently through 2022.

This selected example illustrates a useful image control. Evaluating changes to unrelated attributes remains part of the disentanglement problem.
2021 course team: Houming Chen, Le Qin, Yutong Bi, Jiaxi Chen, Dongyang Zhao.