Meta Muse Image: The Backlash Over Training on Your Photos

Short version: Meta shipped an image generator called Muse Image, and the story is not the model. It is that people objected to their own uploaded photos being used to train it. It reached fal's catalogue on 26 August 2026 with no published price, which is why it is not on our picker.

Disclosure: Leaxor is mine. The last section explains a decision not to sell something, which is the opposite of the usual bias, but the bias is worth flagging either way.

What Meta shipped

Muse Image is Meta's text-to-image model, rolled out from July 2026 inside its own apps rather than as a standalone tool. A companion, Muse Spark 1.2, followed on 6 August.

Distribution is the interesting part of the strategy. Meta does not need people to seek out an image generator; it can put one inside products that billions already open daily. That is a structural advantage no independent lab can match, and it is roughly the same play Google is running by putting Nano Banana into Search and the Gemini app.

The objection

TechCrunch's launch coverage led with the pushback rather than the product: users objecting to the use of their photos.

The complaint is worth stating precisely, because it is easy to flatten into a general anxiety about AI. It is not that Meta trained a model. It is that the material came from people's own uploaded photo libraries, under terms they had accepted years earlier for a different purpose, and that no meaningful moment of consent existed for this particular use.

Whether that is legal and whether it is reasonable are separate questions, and the answers can differ. A terms-of-service update from 2019 is legally load-bearing and practically meaningless as a statement of what someone agreed to.

My own read, flagged as opinion: the objection is legitimate and will keep recurring, because the pattern is not specific to Meta. Any platform holding a decade of user uploads is sitting on the most valuable training set it will ever have, and the temptation to use it does not go away.

Why it is not on our picker

This one is simpler than the ethics, and it is the reason that actually decided it.

Muse Image appeared in fal's catalogue on 26 August 2026 with no published price. Every other model we sell has a rate we can read, multiply, and turn into a credit cost.

An endpoint with an unknown cost cannot be resold responsibly. We would be charging you a fixed number of credits against an upstream bill we could not calculate, which is either a hidden loss or a hidden markup and we would not know which until the invoice arrived. So it stays off until fal publishes a rate.

That is a boring reason, and it is the real one. The training-data question is a fair thing to weigh separately, but it is not what made the call here.

What to use instead

If what appealed was a fast, cheap general-purpose image model, the field is well covered:

• Nano Banana 2 at 1 credit. Google's Flash tier, quick and close to Pro quality.
• Ideogram v4 at 1 credit. The pick when the image contains text.
• Flux 2 Pro at 1 credit. The cheapest good option, and our default.

All are in Leaxor's image generator, priced per image with credits at $0.10 and no subscription.

The part worth taking away

Muse Image will probably be used more than most models on this list, because it is where the people already are. Reach and quality are different axes, and distribution usually wins.

The consent question is the one to keep watching. It is not really about Meta, and it will not stay confined to image generation.

The rest of this month's launches, including the ones we did add, are in the August 2026 roundup.

Reading this with an AI assistant? Ask it to summarize this in ChatGPT, Claude, or Perplexity.