Unlocking Dataset Distillation with Diffusion Models

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Moser, Brian B., Raue, Federico, Palacio, Sebastian, Frolov, Stanislav, Dengel, Andreas
Natura: Preprint
Pubblicazione: 2024
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866916995044212736
author Moser, Brian B.
Raue, Federico
Palacio, Sebastian
Frolov, Stanislav
Dengel, Andreas
author_facet Moser, Brian B.
Raue, Federico
Palacio, Sebastian
Frolov, Stanislav
Dengel, Andreas
contents Dataset distillation seeks to condense datasets into smaller but highly representative synthetic samples. While diffusion models now lead all generative benchmarks, current distillation methods avoid them and rely instead on GANs or autoencoders, or, at best, sampling from a fixed diffusion prior. This trend arises because naive backpropagation through the long denoising chain leads to vanishing gradients, which prevents effective synthetic sample optimization. To address this limitation, we introduce Latent Dataset Distillation with Diffusion Models (LD3M), the first method to learn gradient-based distilled latents and class embeddings end-to-end through a pre-trained latent diffusion model. A linearly decaying skip connection, injected from the initial noisy state into every reverse step, preserves the gradient signal across dozens of timesteps without requiring diffusion weight fine-tuning. Across multiple ImageNet subsets at 128x128 and 256x256, LD3M improves downstream accuracy by up to 4.8 percentage points (1 IPC) and 4.2 points (10 IPC) over the prior state-of-the-art. The code for LD3M is provided at https://github.com/Brian-Moser/prune_and_distill.
format Preprint
id arxiv_https___arxiv_org_abs_2403_03881
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Unlocking Dataset Distillation with Diffusion Models
Moser, Brian B.
Raue, Federico
Palacio, Sebastian
Frolov, Stanislav
Dengel, Andreas
Computer Vision and Pattern Recognition
Artificial Intelligence
Machine Learning
Dataset distillation seeks to condense datasets into smaller but highly representative synthetic samples. While diffusion models now lead all generative benchmarks, current distillation methods avoid them and rely instead on GANs or autoencoders, or, at best, sampling from a fixed diffusion prior. This trend arises because naive backpropagation through the long denoising chain leads to vanishing gradients, which prevents effective synthetic sample optimization. To address this limitation, we introduce Latent Dataset Distillation with Diffusion Models (LD3M), the first method to learn gradient-based distilled latents and class embeddings end-to-end through a pre-trained latent diffusion model. A linearly decaying skip connection, injected from the initial noisy state into every reverse step, preserves the gradient signal across dozens of timesteps without requiring diffusion weight fine-tuning. Across multiple ImageNet subsets at 128x128 and 256x256, LD3M improves downstream accuracy by up to 4.8 percentage points (1 IPC) and 4.2 points (10 IPC) over the prior state-of-the-art. The code for LD3M is provided at https://github.com/Brian-Moser/prune_and_distill.
title Unlocking Dataset Distillation with Diffusion Models
topic Computer Vision and Pattern Recognition
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2403.03881