SeaArt.AI Joins Hands with the PixArt-α Model of Huawei Noah's Ark Lab

SeaArt.AI collaborates with the PixArt-α model of Huawei Noah's Ark Lab to provide smoother and more innovative user experience and initiate a new epoch of AI art creation!

About SeaArt

SeaArt.AI is a technology company that takes a global lead in the AI Generated Content field, focusing on the R&D and application of Generative AI technologies. Its core product, SeaArt, is an easy-to-use, efficient, and user-friendly AI painting tool.

About Huawei Noah's Ark Lab

Huawei Noah's Ark Lab is part of Huawei's head research group "2012 Lab". Founded in 2012, Huawei Noah's Ark Lab, the AI research center of Huawei, is dedicated to researching AI basic algorithms and building an AI engine with high data and energy efficiency.

SeaArt.AI builds a partnership with Huawei Noah's Ark Lab to facilitate the development of the PixArt-α model. This technology focuses on image and art creation, aimed at using AI algorithms to promote a new way of expressing creativity and art.

About PixArt-α

PixArt-α is a T2l diffusion model based on the Transformer. Its generated image quality is comparable to that of the most advanced image generators (Imagen, SDXL, even Midjourney, etc.), and its training speed is significantly higher than the current popular image generators.

The training time of PixArt-α is only 10.8% of that of Stable Diffusion V1.5 (A100 GPU: about 675 days vs 6250 days), saving around 300 thousand USD (26,000 USD vs 320,000 USD) and reducing 90% of CO2 emissions. Besides, compared with the greater SOTA model RAPHAEL, the training costs are only 1%.

To achieve the goal, PixArt-α brings about three core designs:

• Training Strategy Split: Design three different training steps to respectively optimize pixel dependency, Txt2Img alignment, and image aesthetic qualities.
• Efficient T2I Transformer: Integrate the cross-attention module into the Diffusion Transformer (DiT) to import the text conditioning and simplify computationally intensive class conditional branches.
• High-information Data: Emphasize the importance of conceptual density to Txt2Img alignment and utilize large-scale vision-language models to automatically annotate dense pseudo captions, assisting in Txt2Img alignment learning.

A plethora of experiments demonstrate that PixArt-α excels in image quality, artistry, and semantic control. The collaboration between SeaArt.AI and PixArt-α has constructed a high-quality and cost-effective generative model. The sharing of knowledge and resources has sparked the fusion of new technologies and creativity, marking a significant step toward building a healthy and vibrant model ecosystem.

This collaboration serves as a paradigm of deep cooperation and the spirit of sharing within the field of artificial intelligence. By merging their unique technological advantages and R&D resources, the two companies propel innovations and applications of AI technology. This collaborative approach aids in driving forward the frontier exploration of AI technology, while also presenting new possibilities for addressing industry challenges and optimizing products and services.

SeaArt.AI will continue to aggregate wisdom and strength from all sides, accelerate technology iterations and upgrades, and continually innovate in the development and application of AI technology, promoting the healthy development of the industry ecosystem.