Scaling Properties of Diffusion Models for Perceptual Tasks

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Ravishankar, Rahul, Patel, Zeeshan, Rajasegaran, Jathushan, Malik, Jitendra
Natura: Preprint
Pubblicazione: 2024
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866917840007725056
author Ravishankar, Rahul
Patel, Zeeshan
Rajasegaran, Jathushan
Malik, Jitendra
author_facet Ravishankar, Rahul
Patel, Zeeshan
Rajasegaran, Jathushan
Malik, Jitendra
contents In this paper, we argue that iterative computation with diffusion models offers a powerful paradigm for not only generation but also visual perception tasks. We unify tasks such as depth estimation, optical flow, and amodal segmentation under the framework of image-to-image translation, and show how diffusion models benefit from scaling training and test-time compute for these perceptual tasks. Through a careful analysis of these scaling properties, we formulate compute-optimal training and inference recipes to scale diffusion models for visual perception tasks. Our models achieve competitive performance to state-of-the-art methods using significantly less data and compute. To access our code and models, see https://scaling-diffusion-perception.github.io .
format Preprint
id arxiv_https___arxiv_org_abs_2411_08034
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Scaling Properties of Diffusion Models for Perceptual Tasks
Ravishankar, Rahul
Patel, Zeeshan
Rajasegaran, Jathushan
Malik, Jitendra
Computer Vision and Pattern Recognition
Artificial Intelligence
In this paper, we argue that iterative computation with diffusion models offers a powerful paradigm for not only generation but also visual perception tasks. We unify tasks such as depth estimation, optical flow, and amodal segmentation under the framework of image-to-image translation, and show how diffusion models benefit from scaling training and test-time compute for these perceptual tasks. Through a careful analysis of these scaling properties, we formulate compute-optimal training and inference recipes to scale diffusion models for visual perception tasks. Our models achieve competitive performance to state-of-the-art methods using significantly less data and compute. To access our code and models, see https://scaling-diffusion-perception.github.io .
title Scaling Properties of Diffusion Models for Perceptual Tasks
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2411.08034