Quantitative ablation study on VAE channel width and multi-level embeddings.

<p>We evaluated three variants using FID and IS metrics: 4-channel VAE (4ch), 8-channel VAE (8ch), and 8ch model with multi-level CLIP embeddings (8ch+ML). Both 8ch and 8ch+ML variants demonstrated improved FID and IS scores compared to 4ch. However, the IS score for 8ch+ML showed a slight dec...

Full description

Saved in:

Bibliographic Details
Main Author:	Mujung Kim (22367223) (author)
Other Authors:	Jisang Yoo (5645609) (author), Soonchul Kwon (1487401) (author), Byung Jun Kim (6433244) (author), Changsik John Pak (22367226) (author), Chong Hyun Won (11986150) (author), Suk-Ho Moon (14784324) (author), Woo Jin Song (17457980) (author), Han Gyu Cha (22367229) (author), Kyung Hee Park (9278105) (author)
Published:	2025
Subjects:	Ecology Immunology Space Science Biological Sciences not elsewhere classified Information Systems not elsewhere classified results showed improvements mitigate detail loss grained detail preservation capture detailed representations synthesize synthetic images quality synthetic images stable diffusion model latent diffusion model conducted classification experiments existing diffusion models based image encoder enhanced data diversity used interpolation techniques real images inpainting models data scarcity classification performance xlink "> trained segmentation reference material proposed technique overall shapes level embeddings generate high employed pre decoder structures clip encoder clinical use approach contributes applied using
Tags:	Add Tag No Tags, Be the first to tag this record!

Quantitative ablation study on VAE channel width and multi-level embeddings.

Similar Items