CosmicMan: A Text-to-Image Foundation Model for Humans

TL;DR

CosmicMan addresses the shortage of high-quality, human-centric text-to-image foundation models by combining a flowing data flywheel (Annotate Anyone) with a pragmatic training framework (Daring) that discretizes dense human descriptions and explicitly aligns attention to body and outfit regions. The CosmicMan-HQ 1.0 dataset (6M high-resolution human images with 115M detailed attributes) enables scalable, accurate learning, while the HOLA loss and data-discretized cross-attention improve fine-grained alignment for dense concepts. Empirical results show superior image fidelity and region-specific text–image alignment versus state-of-the-art models, with strong performance in 2D editing and 3D reconstruction—demonstrating practical impact for human-centric content generation. The work also emphasizes a sustainable data-and-model governance plan, including ongoing dataset updates and platform-level commitments to privacy and accessibility.

Abstract

We present CosmicMan, a text-to-image foundation model specialized for generating high-fidelity human images. Unlike current general-purpose foundation models that are stuck in the dilemma of inferior quality and text-image misalignment for humans, CosmicMan enables generating photo-realistic human images with meticulous appearance, reasonable structure, and precise text-image alignment with detailed dense descriptions. At the heart of CosmicMan's success are the new reflections and perspectives on data and models: (1) We found that data quality and a scalable data production flow are essential for the final results from trained models. Hence, we propose a new data production paradigm, Annotate Anyone, which serves as a perpetual data flywheel to produce high-quality data with accurate yet cost-effective annotations over time. Based on this, we constructed a large-scale dataset, CosmicMan-HQ 1.0, with 6 Million high-quality real-world human images in a mean resolution of 1488x1255, and attached with precise text annotations deriving from 115 Million attributes in diverse granularities. (2) We argue that a text-to-image foundation model specialized for humans must be pragmatic -- easy to integrate into down-streaming tasks while effective in producing high-quality human images. Hence, we propose to model the relationship between dense text descriptions and image pixels in a decomposed manner, and present Decomposed-Attention-Refocusing (Daring) training framework. It seamlessly decomposes the cross-attention features in existing text-to-image diffusion model, and enforces attention refocusing without adding extra modules. Through Daring, we show that explicitly discretizing continuous text space into several basic groups that align with human body structure is the key to tackling the misalignment problem in a breeze.

Figures (23)

• Figure 1: Data and Annotation Samplings. Sample images with captions obtained from different methods: the original text uploaded along with the images, captions generated by BLIP blip, and InstructBLIP instructblip pretrained models, and our attribute-based text. Here we use different colors to highlight the unrelated, wrong, and coarse and vague descriptions.
• Figure 2: Data Production Paradigm. (a) Data production by humans and (b) data production by AI. (c) Our proposed new data production paradigm by Human-AI cooperation, named Annotate Anyone. It serves as a data flywheel to produce dynamic up-to-date high-quality data at a low cost.
• Figure 2: Effectiveness of Annotate Anyone. The bar charts represent the manual annotation counts for different attributes in $1000$ images for each iteration, and the solid lines depict the accuracy of each attribute under the current model. Better zoom in for details.
• Figure 3: Parsing Examples from CosmicMan-HQ. The parsing results of sampled image in our dataset, along with detailed labels for each part. Text descriptions are obtained from labels.
• Figure 3: CosmicMan-HQ 1.0 Statistics. The left part shows the global attributes and garment types of the person in the image, and the right section displays attributes examples of top garments.