Baidu says Ernie's text-to-image model was trained on publicly available data worldwide

What happened

Days after Ernie Bot's March 2023 launch, users noted its text-to-image output confused the country Turkey with the bird and a crane (the bird) with a crane (the machine), and art influencer Liu Dake publicly questioned whether Baidu was simply translating Chinese prompts into English and routing them through foreign image generators such as Stable Diffusion.

What Baidu said

Baidu said Ernie Bot was 'totally self-developed' and that its text-to-image capability was trained by its Ernie-ViLG model, in a post on its official Weibo account. Baidu also clarified that it uses 'publicly available data worldwide to train the model, which is common industry practice'.

What this establishes — and what it does not

This establishes Baidu's own account of its training-data sourcing: publicly available data worldwide. It does not establish that any specific creator's work was in the training set, and the allegation that Ernie copied a foreign generator is an allegation Baidu denied, not a finding.