Meta's new AI image generator was trained on 1.1B Instagram and FB photos

Availability, Region Locking, and Onboarding

• Many users report “Not available in your location” across New Zealand, EU countries (Netherlands, Norway, Switzerland), Canada, Brazil, India, etc.
• Some speculate about GDPR/EU AI Act or other regional laws, but the real reason is unclear.
• The site often forces users through creating/merging a Meta account (Facebook, Instagram, Oculus/Horizon) before finally saying the service isn’t available, which is widely criticized as manipulative and confusing.
• Access appears to be IP-geolocation based; even accounts set to US can be blocked via non-US VPNs.

Model Quality, UX, and Usefulness

• Those who got access say it’s very fast and roughly comparable to SDXL in quality, behind Midjourney and DALL·E 3 for many prompts.
• It struggles notably with fingers and text rendering; some “hard prompts” (object-in-object, unusual compositions) fail similarly to other models.
• Strong guardrails: it refuses explicit nudity and well-known real people by name, though indirect prompts can sometimes yield celebrity lookalikes.

Alternatives and Moderation Trade-offs

• Several commenters argue that heavily moderated, locked-down generators are less attractive than local Stable Diffusion setups (Auto1111, ComfyUI, Fooocus, etc.), which are more flexible and largely uncensored.
• Debate over which SDXL wrappers are best; opinions vary widely by use case and preferred complexity.

Training Data, Copyright, and Terms of Service

• Long debate on whether training on user-generated Instagram/Facebook content is “copyright laundering.”
• Instagram/Facebook ToS grant Meta broad, worldwide, sublicensable rights to use and create derivative works from uploaded content, but not all agree that users realize this includes AI training.
• Some argue training is likely lawful fair use and not a derivative work; others say this is a legal gray area and a “bait-and-switch” relative to the original purpose of uploads.
• Questions arise about deletion: if a user deletes a photo, does Meta need to remove its influence from models? This is acknowledged as unresolved and potentially very expensive.

Liability and Infringing Outputs

• Distinction is drawn between training data legality and whether specific generated images infringe copyright or trademarks (e.g., Mickey Mouse, Iron Man).
• Many expect liability to fall primarily on the user who publishes infringing outputs, analogous to using a photocopier or Photoshop, though this is not yet firmly settled in courts.

Privacy, Ethics, and User Trust

• Some see this as “yet another reason” to delete Meta accounts; others note Meta has already done worse, so this is incremental.
• Concerns extend beyond copyright to rights of people depicted in photos uploaded by others, and to the idea that “images of ourselves” are now embedded in a commercial AI.

Deepfakes, Nudes, and Harm

• Several comments predict or note existing tools that generate nude or compromising images of real people; others point out Meta’s model is currently heavily nerfed in that direction.
• There is disagreement on how new or harmful this is: some say such misuse predates modern AI but acknowledge that greater scale and ease have increased real-world damage, especially to teenagers and in politics.
• A minority argue that widespread availability of fake nudes might eventually desensitize society, while others emphasize ongoing, serious harm.