Toffee-5M
收藏资源简介:
Toffee-5M是由Adobe Research和加州大学圣克鲁兹分校合作创建的大型数据集,专为特定主题驱动的文本到图像生成和编辑任务设计。该数据集包含480万对图像,其中包括160万对编辑图像,涵盖了风格、背景、颜色等多种变化。创建过程中,利用了预训练的扩散模型和控制网络,无需对每个主题进行微调,显著降低了构建大规模数据集的计算成本。Toffee-5M的应用领域广泛,旨在通过零样本学习,实现对任意主题图像的快速定制生成和编辑,无需测试时微调,极大地推动了相关领域的研究进展。
Toffee-5M is a large-scale dataset co-developed by Adobe Research and the University of California, Santa Cruz, specifically tailored for subject-driven text-to-image generation and editing tasks. This dataset contains 4.8 million image pairs, including 1.6 million edited image pairs, covering diverse variations such as style, background, and color. During its creation, pre-trained diffusion models and ControlNets were utilized, eliminating the need for fine-tuning on each individual subject, which significantly reduced the computational cost of building such a large-scale dataset. Toffee-5M has broad application prospects, aiming to enable rapid custom generation and editing of images for any subject through zero-shot learning without requiring fine-tuning during inference, greatly advancing research progress in related fields.

- 1Toffee: Efficient Million-Scale Dataset Construction for Subject-Driven Text-to-Image GenerationAdobe Research 加州大学圣克鲁兹分校 · 2024年



