Audio-based generative models for music have seen great strides recently, but
so far have not managed to produce full-length music tracks with coherent
musical structure. We show that by training a generative model on long temporal
contexts it is possible to produce long-form music of up to 4m45s. Our model
consists of a diffusion-transformer operating on a highly downsampled
continuous latent representation (latent rate of 21.5Hz). It obtains
state-of-the-art generations according to metrics on audio quality and prompt
alignment, and subjective tests reveal that it produces full-length music with
coherent structure.

通过对长时态上下文的生成模型进行训练，我们展示了可以产生长达 4 分 45 秒的音乐的可能性。我们的模型是在高度降采样的连续潜在表示（潜在速率为 21.5Hz）上操作的扩散变换器，根据音频质量和提示对齐的指标获得了最先进的生成结果，主观测试表明它能产生具有连贯结构的完整音乐。

潜在扩散的长篇音乐生成

Long-form music generation with latent diffusion

As synthetic data becomes higher quality and proliferates on the internet,
machine learning models are increasingly trained on a mix of human- and
machine-generated data. Despite the successful stories of using synthetic data
for representation learning, using synthetic data for generative model training
creates "self-consuming loops" which may lead to training instability or even
collapse, unless certain conditions are met. Our paper aims to stabilize
self-consuming generative model training. Our theoretical results demonstrate
that by introducing an idealized correction function, which maps a data point
to be more likely under the true data distribution, self-consuming loops can be
made exponentially more stable. We then propose self-correction functions,
which rely on expert knowledge (e.g. the laws of physics programmed in a
simulator), and aim to approximate the idealized corrector automatically and at
scale. We empirically validate the effectiveness of self-correcting
self-consuming loops on the challenging human motion synthesis task, and
observe that it successfully avoids model collapse, even when the ratio of
synthetic data to real data is as high as 100%.

通过引入理想化的校正函数，本文旨在稳定自消耗生成模型的训练，理论结果证明了根据真实数据分布将数据点映射为更可能的方法可以使自消耗循环的稳定性呈指数级增加，我们还提出了自我校正函数，它依赖于专家知识并旨在在规模上自动近似理想化的校正器，在具有挑战性的人体运动合成任务上对自我校正自消耗循环的有效性进行了实证验证，观察到即使合成数据与真实数据的比例高达 100%，也能成功避免模型崩溃。

用于生成模型训练的自校正自消耗循环

Self-Correcting Self-Consuming Loops for Generative Model Training

Exploring generative model training for synthetic tabular data, specifically
in sequential contexts such as credit card transaction data, presents
significant challenges. This paper addresses these challenges, focusing on
attaining both high fidelity to actual data and optimal utility for machine
learning tasks. We introduce five pre-processing schemas to enhance the
training of the Conditional Probabilistic Auto-Regressive Model (CPAR),
demonstrating incremental improvements in the synthetic data's fidelity and
utility. Upon achieving satisfactory fidelity levels, our attention shifts to
training fraud detection models tailored for time-series data, evaluating the
utility of the synthetic data. Our findings offer valuable insights and
practical guidelines for synthetic data practitioners in the finance sector,
transitioning from real to synthetic datasets for training purposes, and
illuminating broader methodologies for synthesizing credit card transaction
time series.

本研究探讨了用于合成表格数据的生成模型训练，特别是在连续上下文中，如信用卡交易数据，它面临着重大挑战。本文解决了这些挑战，重点在于实现数据的高保真度和机器学习任务的最佳效用。我们引入了五个预处理方案来增强条件概率自回归模型 (CPAR) 的训练，展示了合成数据保真度和效用的渐进改进。在达到令人满意的保真度水平后，我们将注意力转向针对时间序列数据进行定制的欺诈检测模型的训练，评估合成数据的效用。我们的发现为金融领域的合成数据实践者提供了有价值的见解和实用指南，从真实数据过渡到合成数据用于训练，并阐明了合成信用卡交易时间序列的更广泛方法。