Bing Image Creator Quality Rollback: Microsoft Responds to User Feedback

Microsoft's Bing Image Creator Faces User Backlash After AI Model Upgrade

Prior to the holiday season, Microsoft announced an upgrade to the artificial intelligence model powering Bing Image Creator, its AI-driven image editing functionality integrated within the Bing search engine.

The company had pledged that the new model – the latest iteration of OpenAI’s DALL-E 3, internally designated PR16 – would enable users to generate images with “higher quality” and at “twice the speed” compared to the previous version.

User Complaints and Model Reversion

However, these promises were not fulfilled. Almost immediately, negative feedback surfaced on platforms like X and Reddit.

One Reddit user expressed disappointment, stating the previously favored DALL-E version was now lost. Another user reported switching to ChatGPT due to the perceived ineffectiveness of Bing Image Creator.

The significant negative response prompted Microsoft to announce a temporary reversion to the prior model while they address the identified problems.

Jordi Ribas, head of search at Microsoft, communicated via X on Tuesday evening that they had successfully replicated some of the reported issues.

He stated a plan to revert to DALL-E 3 PR13 until a resolution is achieved. The rollout of this change is a lengthy process, initiated over a week ago and projected to take an additional 2-3 weeks for complete implementation.

What Caused the Issues?

Evaluating model performance based on user reports can be complex, especially without standardized prompts for comparison.

Many users reported that images generated by PR16 appeared less realistic overall. Mayank Parmar, writing for Windows Latest, observed a lack of detail and refinement in PR16 outputs.

The images were described as strangely cartoonish and lacking vibrancy.

Past Instances of AI Model Issues

This isn't an isolated incident. Previously, Google was compelled to temporarily suspend its Gemini AI chatbot’s image creation capabilities.

This pause occurred after users highlighted inaccuracies in the historical depictions generated by the AI.

Challenges in Measuring AI Model Improvement

These events underscore the difficulties inherent in accurately gauging improvements in AI models within real-world applications.

According to Ribas, Microsoft’s internal benchmarks indicated that PR16’s quality was “a bit better on average” than the preceding Bing Image Creator model.

Despite these internal findings, it’s evident that the company’s metrics did not align with the preferences of the majority of users.

Stay informed with TechCrunch’s AI newsletter! Subscribe here to receive it weekly on Wednesdays.