Presentation / Installation


CLIP-PAE: Projection-Augmentation Embedding to Extract Relevant Features for a Disentangled, Interpretable and Controllable Text-Guided Face Manipulation
SessionText-Guided Generation
DescriptionWe found that Contrastive Language-Image Pre-Training (CLIP) embeds texts and images in different regions in the joint space, leading to artifacts in the resulting images when optimizing toward text embeddings. Disentanglement, interpretability, and controllability are hard to guarantee. Therefore, we introduce CLIP projection-augmentation embedding (PAE) as an alternative optimization target.
Authors
Chenliang Zhou
University of Cambridge
Fangcheng Zhong
University of Cambridge
Cengiz Oztireli
University of Cambridge
Google Research
Event Type
Technical Paper
TimeThursday, 10 August 20239:41am - 9:51am PDT
LocationPetree Hall D
ACM Digital Library Technical Paper PDF
Session Time & Location
Sunday, 6 August 20236pm - 8:30pm PDTWest Hall B
Thursday, 10 August 20239am - 10:30am PDTPetree Hall D
Interest Areas
Research & Education
Recordings
Livestreamed
Recorded
Keywords
Artificial Intelligence/Machine Learning
Video
Registration Categories
FC
FCS
V
VS
EFC
Next PresentationNext Presentation
Text-Guided Generation - Q&A and Interactive Discussion