No Tiananmen Square in ERNIE-ViLG, the new Chinese image-making AI

When a demo of the software was released in late August, users quickly noticed that certain words — both explicit mentions of political leaders’ names and potentially controversial words. debate only in a political context — labeled as “sensitive” and blocked from producing any results. It looks like China’s sophisticated online censorship system has expanded into the latest trend in AI.

It’s not uncommon for similar AIs to restrict users from creating certain types of content. DALL-E 2 prohibit sexual content, public figure’s face, or medical treatment image. But the case of ERNIE-ViLG highlights the question of where exactly the line between censorship and political censorship lies.

The ERNIE-ViLG model is part of Wenxin, a large-scale natural language processing project by China’s leading AI company, Baidu. It was trained on a dataset of 145 million image-text pairs and contains 10 billion parameters — values ​​that the neural network adjusts as it learns, which the AI ​​uses to distinguish subtle differences. between concepts and art styles.

That means ERNIE-ViLG has a smaller training dataset than DALL-E 2 (650 million pairs) and Diffuse Stable (2.3 billion pairs) but more parameters than either (DALL-E 2 has 3.5 billion parameters and Steady Diffusion has 890 million). Baidu released a demo version on its own platform at the end of August and then after huggingPopular international AI community.

The main difference between ERNIE-ViLG and the Western models is that the Baidu-developed model understands prompts written in Chinese and is less likely to make mistakes when it comes to culture-specific words.

For example, a Chinese video creator compared results from different models to find prompts that included Chinese historical figures, popular culture and food celebrities. He found that ERNIE-ViLG produced more accurate images than DALL-E 2 or Diffuse Stabilization. After its release, ERNIE-ViLG was also approved by the insiders Japanese anime communitywho have found that this model can produce more satisfying anime art than other models, possibly because it includes more anime in its training data.

But ERNIE-ViLG will be defined, like other models, according to what it allows. Unlike DALL-E 2 or Stable Diffusion, ERNIE-ViLG has no published explanation of its content moderation policy, and Baidu declined to comment for this story.

When the ERNIE-ViLG demo was first released on Hugging Face, users who entered certain words received the message “Sensitive words found. Please type again (存在 敏感 词 , 请 重新 输入), ” that is a surprisingly honest admission of the filtering mechanism. However, as of at least September 12, the message reads “The imported content does not meet the relevant rules. Please try again after adjusting. (输入 ) ”

Source link


Kig News: Update the world's latest breaking news online of the day, breaking news, politics, society today, international mainstream news .Updated news 24/7: Entertainment, the World everyday world. Hot news, images, video clips that are updated quickly and reliably

Related Articles

Back to top button