ViTok-v2 scales to 5B parameters, advancing image autoencoder reconstruction and generation

Researchers have introduced ViTok-v2, a 5-billion parameter image autoencoder that scales to larger resolutions and parameter counts than previous models. This new model utilizes native resolution support and a DINOv3 perceptual loss to achieve better reconstruction quality across various image sizes. ViTok-v2 was trained on approximately 2 billion images and demonstrates improved performance at higher resolutions compared to existing methods. AI

影响 Advances the state-of-the-art in image autoencoders, potentially improving generative model capabilities.

排序理由 This is a research paper detailing a new model architecture and its performance. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.LG 阅读 →

• arXiv
• DINOv3
• GAN
• Hansen-Estruch
• LPIPS
• ViTok-v2

• 论文
• 模型发布

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →