Our work focuses on addressing sample deficiency from low-density regions of
data manifold in common image datasets. We leverage diffusion process based
generative models to synthesize novel images from low-density regions. We
observe that uniform sampling from diffusion models predominantly samples from
high-density regions of the data manifold. Therefore, we modify the sampling
process to guide it towards low-density regions while simultaneously
maintaining the fidelity of synthetic data. We rigorously demonstrate that our
process successfully generates novel high fidelity samples from low-density
regions. We further examine generated samples and show that the model does not
memorize low-density data and indeed learns to generate novel samples from
low-density regions.

我们使用扩散过程生成模型解决了常见图像数据集中低密度区域的样本稀缺问题，通过修改采样过程使其指向低密度区域并同时保持合成数据的逼真度，经过严谨的验证，我们的方法成功地生成了低密度区域的新的高保真度样本。

使用扩散模型从低密度区域生成高保真数据

Generating High Fidelity Data from Low-density Regions using Diffusion  Models

The objective of this work is to reconstruct an original image from
Bag-of-Visual-Words (BoVW). Image reconstruction from features can be a means
of identifying the characteristics of features. Additionally, it enables us to
generate novel images via features. Although BoVW is the de facto standard
feature for image recognition and retrieval, successful image reconstruction
from BoVW has not been reported yet. What complicates this task is that BoVW
lacks the spatial information for including visual words. As described in this
paper, to estimate an original arrangement, we propose an evaluation function
that incorporates the naturalness of local adjacency and the global position,
with a method to obtain related parameters using an external image database. To
evaluate the performance of our method, we reconstruct images of objects of 101
kinds. Additionally, we apply our method to analyze object classifiers and to
generate novel images via BoVW.

本研究旨在从视觉单词（BoVW）中重建原始图像，提出了一种含自然性本地邻近性及全局位置的评估函数，并通过使用外部图像数据库获取相关参数的方法来估计原始布局，以达到成功重建 BoVW 的目的，并产生了 101 种不同物体的图像和新颖的 BoVW 图像。