Analyzing and Mitigating Model Collapse in Rectified Flow Models

• URL: http://arxiv.org/abs/2412.08175v2
• Date: Sun, 09 Feb 2025 10:02:55 GMT
• Title: Analyzing and Mitigating Model Collapse in Rectified Flow Models
• Authors: Huminhao Zhu, Fangyikang Wang, Tianyu Ding, Qing Qu, Zhihui Zhu,
• Abstract summary: Recent studies have shown that repeatedly training on self-generated samples can lead to model collapse. We provide both theoretical analysis and practical solutions for addressing MC in diffusion/flow models. We propose a novel Real-data Augmented Reflow and a series of improved variants, which seamlessly integrate real data into Reflow training by leveraging reverse flow.
• Score: 23.568835948164065
• License:
• Abstract: Training with synthetic data is becoming increasingly inevitable as synthetic content proliferates across the web, driven by the remarkable performance of recent deep generative models. This reliance on synthetic data can also be intentional, as seen in Rectified Flow models, whose Reflow method iteratively uses self-generated data to straighten the flow and improve sampling efficiency. However, recent studies have shown that repeatedly training on self-generated samples can lead to model collapse (MC), where performance degrades over time. Despite this, most recent work on MC either focuses on empirical observations or analyzes regression problems and maximum likelihood objectives, leaving a rigorous theoretical analysis of reflow methods unexplored. In this paper, we aim to fill this gap by providing both theoretical analysis and practical solutions for addressing MC in diffusion/flow models. We begin by studying Denoising Autoencoders and prove performance degradation when DAEs are iteratively trained on their own outputs. To the best of our knowledge, we are the first to rigorously analyze model collapse in DAEs and, by extension, in diffusion models and Rectified Flow. Our analysis and experiments demonstrate that rectified flow also suffers from MC, leading to potential performance degradation in each reflow step. Additionally, we prove that incorporating real data can prevent MC during recursive DAE training, supporting the recent trend of using real data as an effective approach for mitigating MC. Building on these insights, we propose a novel Real-data Augmented Reflow and a series of improved variants, which seamlessly integrate real data into Reflow training by leveraging reverse flow. Empirical evaluations on standard image benchmarks confirm that RA Reflow effectively mitigates model collapse, preserving high-quality sample generation even with fewer sampling steps.

Related papers

• Feasible Learning [78.6167929413604]
  We introduce Feasible Learning (FL), a sample-centric learning paradigm where models are trained by solving a feasibility problem that bounds the loss for each training sample. Our empirical analysis, spanning image classification, age regression, and preference optimization in large language models, demonstrates that models trained via FL can learn from data while displaying improved tail behavior compared to ERM, with only a marginal impact on average performance.
  arXiv  Detail & Related papers  (2025-01-24T20:39:38Z)
• FlowDAS: A Flow-Based Framework for Data Assimilation [15.64941169350615]
  FlowDAS is a novel generative model-based framework using the interpolants to unify the learning of state transition dynamics and generative priors. Our experiments demonstrate FlowDAS's superior performance on various benchmarks, from the Lorenz system to high-dimensional fluid superresolution tasks.
  arXiv  Detail & Related papers  (2025-01-13T05:03:41Z)
• E2EDiff: Direct Mapping from Noise to Data for Enhanced Diffusion Models [15.270657838960114]
  Diffusion models have emerged as a powerful framework for generative modeling, achieving state-of-the-art performance across various tasks. They face several inherent limitations, including a training-sampling gap, information leakage in the progressive noising process, and the inability to incorporate advanced loss functions like perceptual and adversarial losses during training. We propose an innovative end-to-end training framework that aligns the training and sampling processes by directly optimizing the final reconstruction output.
  arXiv  Detail & Related papers  (2024-12-30T16:06:31Z)
• Towards Theoretical Understandings of Self-Consuming Generative Models [56.84592466204185]
  This paper tackles the emerging challenge of training generative models within a self-consuming loop. We construct a theoretical framework to rigorously evaluate how this training procedure impacts the data distributions learned by future models. We present results for kernel density estimation, delivering nuanced insights such as the impact of mixed data training on error propagation.
  arXiv  Detail & Related papers  (2024-02-19T02:08:09Z)
• Guided Flows for Generative Modeling and Decision Making [55.42634941614435]
  We show that Guided Flows significantly improves the sample quality in conditional image generation and zero-shot text synthesis-to-speech. Notably, we are first to apply flow models for plan generation in the offline reinforcement learning setting ax speedup in compared to diffusion models.
  arXiv  Detail & Related papers  (2023-11-22T15:07:59Z)
• Exploiting Diffusion Prior for Real-World Image Super-Resolution [75.5898357277047]
  We present a novel approach to leverage prior knowledge encapsulated in pre-trained text-to-image diffusion models for blind super-resolution. By employing our time-aware encoder, we can achieve promising restoration results without altering the pre-trained synthesis model.
  arXiv  Detail & Related papers  (2023-05-11T17:55:25Z)
• Exploring Continual Learning of Diffusion Models [24.061072903897664]
  We evaluate the continual learning (CL) properties of diffusion models. We provide insights into the dynamics of forgetting, which exhibit diverse behavior across diffusion timesteps.
  arXiv  Detail & Related papers  (2023-03-27T15:52:14Z)
• A Physics-informed Diffusion Model for High-fidelity Flow Field Reconstruction [0.0]
  We propose a diffusion model which only uses high-fidelity data at training. With different configurations, our model is able to reconstruct high-fidelity data from either a regular low-fidelity sample or a sparsely measured sample. Our model can produce accurate reconstruction results for 2d turbulent flows based on different input sources without retraining.
  arXiv  Detail & Related papers  (2022-11-26T23:14:18Z)
• Contrastive Model Inversion for Data-Free Knowledge Distillation [60.08025054715192]
  We propose Contrastive Model Inversion, where the data diversity is explicitly modeled as an optimizable objective. Our main observation is that, under the constraint of the same amount of data, higher data diversity usually indicates stronger instance discrimination. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate that CMI achieves significantly superior performance when the generated data are used for knowledge distillation.
  arXiv  Detail & Related papers  (2021-05-18T15:13:00Z)
• DeFlow: Learning Complex Image Degradations from Unpaired Data with Conditional Flows [145.83812019515818]
  We propose DeFlow, a method for learning image degradations from unpaired data. We model the degradation process in the latent space of a shared flow-decoder network. We validate our DeFlow formulation on the task of joint image restoration and super-resolution.
  arXiv  Detail & Related papers  (2021-01-14T18:58:01Z)
• Model-based Policy Optimization with Unsupervised Model Adaptation [37.09948645461043]
  We investigate how to bridge the gap between real and simulated data due to inaccurate model estimation for better policy optimization. We propose a novel model-based reinforcement learning framework AMPO, which introduces unsupervised model adaptation. Our approach achieves state-of-the-art performance in terms of sample efficiency on a range of continuous control benchmark tasks.
  arXiv  Detail & Related papers  (2020-10-19T14:19:42Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.