谢赛宁转述 RAE 在真实 T2I 规模上的验证结果并开源代码模型数据
> "rae can’t scale"
谢赛宁转述 Peter Tong 团队的进展：此前有人质疑 Representation Autoencoders（RAE）无法扩展、无法泛化到 ImageNet 之外、做不好细节，团队没有在网上争论，而是直接在真实 T2I 规模上尝试，结果看起来非常乐观。他同时致谢 Peter、Boyang、Austin 及所有参与者，并表示代码、模型和数据全部开源。Peter Tong 原帖提到，去年 10 月他们提出 RAE，证明在冻结语义表征上训练扩散模型可行，并在 ImageNet 上优于 VAE，此次回答了它能否扩展到 T2I 等复杂场景的问题。
查看原帖Last October, we introduced Representation Autoencoders (RAE), showing that training diffusion on frozen semantic representations works and outperforms VAEs on ImageNet. We received many questions: Can this scale to complex settings like T2I? Do the advantages hold? The answer is YES. 🧵
正文
当前显示原文，中文全文译文尚未就绪。
> "rae can’t scale" > "rae can’t generalize past imagenet" > "rae can’t do details"
> instead of arguing online > students put heads down > try it at real t2i scale > results come back > look extremely bullish
> shoutout to peter, boyang, austin > and everyone who shipped
> code, model, data > all open-sourced 👇