From SRA to Self-Flow: Data Augmentation or Self-Supervision?

• 类型：arxiv
• 标识：2607.02508
• 链接：https://arxiv.org/abs/2607.02508
• 主分类：multimodal
• 形态：method
• 被引：0
• 被引来源：Semantic Scholar + OpenAlex
• S2被引：0
• OpenAlex被引：0
• 影响力被引：0
• TLDR：Attention Separation is introduced, which preserves the same dual-timestep input as Self-Flow while blocking attention between tokens assigned to different noise levels, and shows that Attention Separation itself provides an augmentation effect by splitting a single image into multiple effective training parts to expand the training data.
• OpenAlex ID：W7167231315
• OpenAlex DOI：10.48550/arxiv.2607.02508
• DOI：10.48550/arxiv.2607.02508
• DOI来源：OpenAlex
• 开放获取：green
• 开放获取链接：https://doi.org/10.48550/arxiv.2607.02508
• OpenAlex更新：2026-07-19
• 副分类：risk
• 待LLM分类：否
• 标题中文：从 SRA 到 Self-Flow：数据增强还是自监督？
• TLDR中文：提出 Attention Separation，在保留与 Self-Flow 相同的双时间步输入的同时，阻止被分配到不同噪声水平 token 之间的注意力，并表明 Attention Separation 本身通过将单张图像拆分为多个有效训练部分来扩充训练数据，从而带来增强效果。
• 来源文件：
• /inbox/tom/_candidates/2026-07-03-agent-rag-longcontext-candidates.json
• [S2 enrich]
• [OpenAlex backfill]