We Put Seven Image Editors in a Blind Lineup. Muse Image Won Our Ballot.

Published: 2026-09-10

We are always working to provide a seamless experience with the best models possible. Image models change fast. The winner last month might not be the winner today. The only way to know is to keep testing.

This time we tested image editing. We had been using Gemini 2.5 Flash Image. Rather than trust a public leaderboard, we put the candidates side by side with the names turned off.

The test

One source photo (a Havana street scene generated in Chat-O) and the same three instructions given to every model:

Night: “Turn this photo into a night scene. Keep the same street, buildings, cars and people, just at night with street lights on.”
No cars: “Remove all the cars from the street. Keep everything else exactly the same.”
Fireworks: “Turn the scene into night and add fireworks in the sky above the street. Keep the buildings, street and people.”

Seven models: Meta Muse Image 1.0, Gemini 3.1 Flash Lite Image, Gemini 2.5 Flash Image, Gemini 3.1 Flash Image, GPT-5 Image mini, FLUX Kontext Pro, and Qwen Image Edit. That is 21 edits.

The test was blind. We hid the names and shuffled the order each round so position would not matter. We cast three blind votes, one per instruction, picking the image we would actually use. We revealed the names only after each round. This is not a massive lab study, and we will not pretend it is.

Each panel below is labeled with the model, so the round is already spoiled for you. Judge the pictures, not the winner.

Night: Muse Image

Night edit, seven models

Muse turned the street to night while keeping the buildings, the cars, and the people exactly where they were. Most of the other models lit the scene warmly. A few panels went so dark the frame is barely readable.

No cars: Muse Image

Remove the cars, seven models

This is the round that separates models. Removing the cars is easy. Leaving everything else alone is the hard part. Muse cleared the street and left the architecture and the passersby in place. Two models left a car in the frame, and others shifted the composition around the edit.

Fireworks: Gemini 2.5 Flash Image

Add fireworks, seven models

Muse lost this round. Gemini 2.5 Flash Image won with the most convincing sky and fireworks. A few others did fine, but GPT-5 Image mini and FLUX Kontext Pro came out way too dark. We are showing the round we lost. A clean sweep would have been harder to believe.

What we took away

Good editing comes down to one rule: change what was asked, and leave the rest alone. Removing the cars should not repaint the buildings. The night edit should not redesign the street. Look back at the no-cars round and you can see which models treated the instruction as an excuse to redraw everything.

Most models fail this simple test. But if version two looks nothing like version one, you cannot build on your work.

Muse won two of the three rounds, so it is the engine we moved image editing to.

It is the default now

Image editing in Chat-O runs on Muse Image through Meta’s Model API. We also moved image generation to Muse, with an automatic fallback to our previous engine if the route is ever down. That was a separate product decision. The comparison above covers editing only.

The model was only half the work. The interface side is in the follow-up post, which goes up tomorrow.

Real photos, too

We love image models, but they do not replace real photography. Plenty of the images people need already exist, taken by someone who was actually standing there.

Here is the same city the AI images above are based on, photographed for real:

Havana, photographed: Plaza de la Catedral

Photo: Tkirkgoz on Pexels.

If you want a real place or a real person, start with Pexels. Free photos from professional photographers, no generation required. We recommend it often, especially when the honest answer is that you do not need an AI at all.

AI is most useful for images that do not exist yet: the scene you described but have never seen, the edit you cannot make with a camera, the version of a photo that only exists in your head.