Microsoft Lens outputs Shutterstock watermarks from unfiltered training data
Lens-Base model generates images with obvious Shutterstock logos intact.
Get AI news that actually matters
One email a day. Zero fluff. Join 10,000+ professionals.
Deep Dive
Lens, trained on a mix of public, licensed, and internal datasets, generated an image (seed 2044664225) that shows the Shutterstock logo in the corner and plastered across the output. The user questions whether the model can detect such watermarks and expresses surprise that watermarked images aren't filtered from the training data.
Key Points
- Lens-Base generates images with Shutterstock watermarks (e.g., corner + overlay) as shown with seed 2044664225.
- Microsoft trained the model on 'public, licensed, and internal datasets' but failed to filter watermarked images.
- This oversight makes outputs unusable for commercial work and highlights weak data curation in AI pipelines.
Why It Matters
Exposes critical training data flaws that could lead to copyright infringement and limit commercial use of AI-generated images.