Isometric Representation Learning for Disentangled Latent Space of Diffusion Models
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: isometric diffusion, disentanglement, diffusion
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: Our method promotes isometry between the latent space and the data manifold, enabling a semantically clear and geometrically sound latent space.
Abstract: Diffusion models have made remarkable progress in capturing and reproducing real-world data. Despite their success and further potential, however, their latent space, the core of diffusion models still remains underexplored. In fact, the latent spaces of existing diffusion models still do not perfectly align with the human perception, entangling multiple concepts in a distorted space. In this paper, we present \textit{Isometric Diffusion}, equipping a diffusion model with isometric representation learning to better reflect human intuition and understanding of visual data. Specifically, we propose a novel loss to promote isometry between the latent space and the data manifold, enabling a semantically clear and geometrically sound latent space. This approach allows smoother interpolation and more precise control over attributes directly in the latent space. Our extensive experiments demonstrate the effectiveness of Isometric Diffusion, marking a significant advance in aligning latent spaces with perceptual semantics. This work paves the way for fine-grained data generation and manipulation.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
Supplementary Material: zip
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1075
Loading