Abstract:We present a novel approach for generating minority samples that live on low-density regions of a data manifold. Our framework is built upon diffusion models, leveraging the principle of guided sampling that incorporates an arbitrary energy-based guidance during inference time. The key defining feature of our sampler lies in its \emph{self-contained} nature, \ie, implementable solely with a pretrained model. This distinguishes our sampler from existing techniques that require expensive additional components (like external classifiers) for minority generation. Specifically, we first estimate the likelihood of features within an intermediate latent sample by evaluating a reconstruction loss w.r.t. its posterior mean. The generation then proceeds with the minimization of the estimated likelihood, thereby encouraging the emergence of minority features in the latent samples of subsequent timesteps. To further improve the performance of our sampler, we provide several time-scheduling techniques that properly manage the influence of guidance over inference steps. Experiments on benchmark real datasets demonstrate that our approach can greatly improve the capability of creating realistic low-likelihood minority instances over the existing techniques without the reliance on costly additional elements. Code is available at \url{<a class="link-external link-https" href="https://github.com/soobin-um/sg-minority" rel="external noopener nofollow">this https URL</a>}.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: **Generate minority samples in the low - density regions of the data manifold**. Specifically, the author focuses on how to generate low - probability data points with unique features using a pre - trained diffusion model without relying on additional expensive components (such as external classifiers). ### Problem Background In modern large - scale datasets, the data distribution is usually a long - tail distribution, which contains a large number of minority samples located in the low - density regions of the data manifold. Although these minority samples are small in number, they often have unique features and are very important for certain tasks (such as classification, anomaly detection, and medical diagnosis). In addition, generating these minority samples can also enhance the predictive ability of the model, promote fairness, and play an important role in creative AI applications. However, under the standard generation frameworks (such as GANs and DDPM), generating a small number of low - probability samples requires a large amount of time and computing resources. Existing methods usually rely on label information or external components (such as classifiers), which are difficult to obtain or costly in practical applications. ### Main Contributions of the Paper To solve the above problems, this paper proposes a self - guided minority - sample generation method, which has the following main features: 1. **Self - guided Sampling**: Self - guidance is achieved through a pre - trained diffusion model without the need for additional external components (such as classifiers), thereby reducing the cost of generating minority samples. 2. **Identification of Low - density Regions**: A new metric is introduced to evaluate the uniqueness of intermediate latent samples. This metric is based on the reconstruction loss and can be directly used during the inference process. 3. **Time Scheduling Technique**: Several time - scheduling methods are provided to rationally manage the influence of the guidance signal and further improve the quality of the generated samples. ### Method Overview The core of this method is to develop a metric that can evaluate low - density regions during the inference process and use it to guide the generation process of the diffusion model. The specific steps are as follows: 1. **Define Low - density Metric**: Evaluate the uniqueness of intermediate latent samples through the reconstruction loss. The formula is: \[ \tilde{L}(x_t; s) := \mathbb{E}_{q_{\alpha_s}(\hat{x}_s | \hat{x}_0)}[d(\hat{x}_0(x_t), \hat{\hat{x}}_0(\hat{x}_s(x_t)))] \] where \(\hat{x}_0(x_t)\) is the denoised version obtained through the Tweedie formula, \(\hat{x}_s(x_t)\) is the sample obtained through the forward - process perturbation, and \(d(\cdot, \cdot)\) is the distance metric. 2. **Self - guided Sampling**: Use the gradient of the above metric as a guidance signal to update the samples in the generation process: \[ g^*(x_t, t; s) := \nabla_{x_t} \mathbb{E}_{q_{\alpha_s}(\hat{x}_s | \hat{x}_0)}[d(\hat{x}_0(x_t), \text{sg}(\hat{\hat{x}}_0(\hat{x}_s(x_t))))] \] where `sg` represents the stop - gradient operation to reduce the computational overhead. 3. **Time Scheduling**: Adjust the time intensity of the guidance signal to avoid conflict with the reverse process and improve the quality of the generated samples. ### Experimental Results Experiments show that this method significantly improves the ability to generate minority samples on multiple benchmark datasets, especially without the need for additional expensive components. The generated samples perform excellently in low - density metrics (such as average k - nearest neighbors and LOF) and quality metrics (such as Fréchet Inception Distance). In conclusion, this paper proposes an efficient and practical method that can generate high - quality minority samples without relying on additional components, which has important theoretical and practical significance.

Self-Guided Generation of Minority Samples Using Diffusion Models

Don't Play Favorites: Minority Guidance for Diffusion Models

Minority-Focused Text-to-Image Generation via Prompt Optimization

Exploring Guided Sampling of Conditional GANs

A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models

AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model

Diffusion Rejection Sampling

Censored Sampling of Diffusion Models Using 3 Minutes of Human Feedback

Towards Faster Non-Asymptotic Convergence for Diffusion-Based Generative Models

Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions

An Overview of Diffusion Models: Applications, Guided Generation, Statistical Rates and Optimization

Fair Sampling in Diffusion Models through Switching Mechanism

Efficient and Controllable Remote Sensing Fake Sample Generation Based on Diffusion Model

Particle Guidance: non-I.I.D. Diverse Sampling with Diffusion Models

Constrained Diffusion with Trust Sampling

A Geometric Perspective on Diffusion Models

Beta Sampling is All You Need: Efficient Image Generation Strategy for Diffusion Models using Stepwise Spectral Analysis

Self-Guided Diffusion Models

Generative adversarial minority enlargement—A local linear over-sampling synthetic method

Correcting Diffusion Generation through Resampling

Accelerating Parallel Sampling of Diffusion Models