论文概要
领域: Computer Vision (CV) 发布时间: 2026-08-18 arXiv: 2608.18076
Summary
Large-scale image generation has benefited from advances in data scale, quality, rebalancing, and recaptioning, yet conventional pipelines typically optimize task-specific datasets in isolation. A central challenge is not only how to curate each task-specific corpus, but also how to organize heterogeneous supervision according to the dependencies among generative capabilities.
The paper presents a capability-driven data infrastructure that couples capability-specific supervision construction with capability-aligned curriculum scheduling.
Key points
- Three specialized data engines build complementary relational supervision for text-image grounding, inter-image transformation, and image-knowledge association; caption experts align T2I and editing supervision across tasks and granularities.
- Multi-stage curriculum co-evolves task composition, visual concept distribution, data quality, and image resolution along the dependency order of capability acquisition.
- Capability-aware evaluation closes the loop via targeted retrieval, expert construction, and gap-aware resampling.
- Scale: the framework curated 440 million T2I images, 120 million editing pairs, and over 27 million image-entity pairs.
- Models: 3B and 6B parameter multimodal diffusion models trained from scratch on this data.
- Results: broad visual coverage, diverse rendering effects, and effective transfer across generative capabilities.
中文摘要
大规模图像生成在数据规模、质量、重平衡和重新标注等方面取得了进展，但传统流程通常孤立地优化特定任务的数据集。核心挑战不仅在于如何策划每个特定任务的语料库，还在于如何根据生成能力之间的依赖关系来组织异构监督。该工作提出能力驱动的数据基础设施，结合三个专业化数据引擎、标题专家对齐与多阶段课程调度，并构建能力感知评估闭环。基于该基础设施策划了4.4亿张文生图图像、1.2亿对编辑数据和超过2700万图像-实体对，从零训练了30亿与60亿参数的多模态扩散模型。
---
*Auto-collected on 2026-08-20.*