LatentCLR: A Contrastive Learning Approach for Unsupervised Discovery of Interpretable Directions

Yüksel, Oğuz Kaan; Simsar, Enis; Er, Ezgi Gülperi; Yanardag, Pinar

doi:10.1109/iccv48922.2021.01400

Cited by 41 publications

(21 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Those vectors may represent interpretable directions. LatentCLR [24] utilise a contrastive learning approach [25] to find usable directions. However, most vectors are invalid or redundant, so the methods should check each vector carefully.…”

Section: Interpretable Directionmentioning

confidence: 99%

“…11). We compared our method with 3 open source methods: Inter-faceGAN [12], GANSpace [15], and LatentCLR [24]. We evaluated 4 attributes: GAN inversion (image reconstruction) and three directions (gender, age, hair).…”

Section: Linear Classifier For W Dmentioning

confidence: 99%

“…Figure 13: Visual comparison among our method, InterfaceGAN [12], GANSpace [15], and LatentCLR [24].…”

Section: Declaration Of Competing Interestmentioning

confidence: 99%

See 2 more Smart Citations

Fast 2-step regularization on style optimization for real face morphing

Cheng¹,

Wang²,

Li³

et al. 2022

Neural Networks

View full text Add to dashboard Cite

Section: Interpretable Directionmentioning

confidence: 99%

Section: Linear Classifier For W Dmentioning

confidence: 99%

See 1 more Smart Citation

Fast 2-step regularization on style optimization for real face morphing

Cheng¹,

Wang²,

Li³

et al. 2022

Neural Networks

View full text Add to dashboard Cite

“…Image Editing Using Latent Space Interpolations. Latent space of pretrained StyleGAN models is highly structured [52] and is popularly used to perform realistic image edits in the generated images [1,4,24,47,52,53,61,65]. The primary idea in most of these approaches is to find a direction in the the extended latent space W+ for editing attributes and transforming a latent code by moving in that direction to perform edits.…”

Section: Related Workmentioning

confidence: 99%

“…Editability. The semantically rich structure of the latent space is widely used for performing semantic edits on the generated images [1,4,49,52,61,65]. For instance, if we have to add the attribute smile to the generated face image, one can edit the latent code as w edit = w + αd where α is edit strength and d is the direction for the smile attribute edit operation.…”

Section: Analysis Of Linearity Of Latent Spacementioning

confidence: 99%

Hierarchical Semantic Regularization of Latent Spaces in StyleGANs

Karmali

Parihar

Agrawal

et al. 2022

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Progress in GANs has enabled the generation of high-resolution photorealistic images of astonishing quality. StyleGANs allow for compelling attribute modification on such images via mathematical operations on the latent style vectors in the W/W+ space that effectively modulate the rich hierarchical representations of the generator. Such operations have recently been generalized beyond mere attribute swapping in the original StyleGAN paper to include interpolations. In spite of many significant improvements in StyleGANs, they are still seen to generate unnatural images. The quality of the generated images is predicated on two assumptions; (a) The richness of the hierarchical representations learnt by the generator, and, (b) The linearity and smoothness of the style spaces. In this work, we propose a Hierarchical Semantic Regularizer (HSR) 1 which aligns the hierarchical representations learnt by the generator to corresponding powerful features learnt by pretrained networks on large amounts of data. HSR is shown to not only improve generator representations but also the linearity and smoothness of the latent style spaces, leading to the generation of more natural-looking style-edited images. To demonstrate improved linearity, we propose a novel metric -Attribute Linearity Score (ALS). A significant reduction in the generation of unnatural images is corroborated by improvement in the Perceptual Path Length (PPL) metric by 16.19% averaged across different standard datasets while simultaneously improving the linearity of attribute-change in the attribute editing tasks.

show abstract