Скачать или смотреть Lumina-Image 2.0: Advanced Text-to-Image Generation

https://arxiv.org/pdf/2503.21758

Lumina-Image 2.0: Advanced Text-to-Image Generation
Lumina-Image 2.0：高级文本到图像生成

Summary: Lumina-Image 2.0 is a new text-to-image (T2I) generative framework that builds upon previous work (Lumina-Next) with a focus on unification and efficiency. It introduces a Unified Next-DiT architecture that processes text and image tokens jointly and a Unified Captioner (UniCap) for generating high-quality, multi-granularity, multi-perspective, and multi-lingual captions. The framework also incorporates multi-stage progressive training and inference acceleration techniques to enhance performance and speed. Lumina-Image 2.0 demonstrates strong results on academic benchmarks and public arenas, even with a relatively modest 2.6B parameters.

摘要：Lumina-Image 2.0 是一个全新的文本转图像 (T2I) 生成框架，它以先前的研究成果 (Lumina-Next) 为基础，注重统一性和效率。它引入了统一的 Next-DiT 架构，用于联合处理文本和图像标记，以及统一的字幕生成器 (UniCap)，用于生成高质量、多粒度、多视角和多语言的字幕。该框架还融合了多阶段渐进式训练和推理加速技术，以提升性能和速度。即使参数规模相对较小，Lumina-Image 2.0 也在学术基准测试和公共领域展现出强劲的成果。