Why AI Merges Two Subjects Into One

AI image models sometimes fuse two subjects into one: you ask for a dog and a cat and get a single animal with features of both, or the same subject twice. This is called concept bleeding, the blurring together of separate concepts from your prompt, and it happens most when the two subjects resemble each other. You keep them apart by giving each subject its own attribute and its own place, naming the relationship between them, keeping your prompt short, and if needed generating them separately and combining. Below you'll read why it happens, when the risk is highest, and how to keep two subjects distinct in your image.

Why does an AI model fuse two subjects?

An AI model spreads its attention across the words in your prompt while generating, and when two subjects land in the same spot in the frame, the model paints their features over each other into a single hybrid. Unlike a dropped subject, nothing disappears here; the two subjects merge instead.

The cause starts with how the text is processed. The researchers behind Isolated Diffusion (2024) describe concept bleeding as the "unexpected overlapping or merging of various concepts", caused by the text encoder compressing your whole prompt into a fixed number of tokens. As a result, separate concepts can interfere during encoding: for "a dog next to a cat" the model may produce two cats, because one concept bleeds over the other.

The study A Cat Is A Cat (Not A Dog!) (NeurIPS 2024) points to a second cause: the information in your prompt biases toward the first-mentioned subject, so features from one subject transfer to the other. The result is a cat that looks dog-like, or two faces that fuse into one average face.

In short: the model won't keep two subjects apart as long as they sit too close together in meaning and in place.

When do two subjects bleed into each other?

The risk is highest when the two subjects resemble each other, because their concepts sit close together for the model. Two animals, two people, or two of the same type of object fuse faster than a person next to a building.

That this is no rare glitch is clear from older benchmark data. In a human evaluation within the VISOR study (Gokhale et al., 2022), DALL·E 2 had the highest share of images with merged objects: 32.46%, nearly one in three images with two subjects. Modern models do this better, but the failure mode hasn't vanished, and it still shows up precisely with lookalike subjects.

You run the most risk with:

• Two subjects from the same category, such as two dogs, two women, or two cars.
• A person plus an animal close together in the frame, like a portrait with a pet on the lap.
• Two faces in one image, which can merge into a single blended face.
• Short, vague labels like "two people" or "a couple of animals", with no attribute per subject.

How to keep two subjects separate in your prompt

The best control is to give each subject its own attribute and its own place, so the model can't stack them in the same spot. You make the two so different and so clearly separated that fusing is no longer an option.

1. Give each subject its own contrasting attribute. "A black labrador and a white persian cat" is harder to fuse than "a dog and a cat", because each subject has its own anchor.
2. Put them in separate places in the frame. Add a spatial split: "a man on the left, a woman on the right" or "the cat in the foreground, the dog in the background". Separate positions force separate subjects.
3. Name the relationship between the two. "A woman walking next to a golden retriever" keeps both apart, because the sentence doesn't hold up without two separate beings.
4. Mention the most important subject first. Because information biases toward the first-mentioned subject, you deliberately give weight to the one that must stay sharpest.
5. Keep your prompt short and focused. The more mood words you stack, the more the concepts bleed together. Cut anything that doesn't help tell the two subjects apart.

Then test small. Generate one image, check that both subjects are separate and recognizable, and adjust your prompt before running a whole series. If you want to build subject, attribute, and place cleanly apart, the prompt generator helps you hold that structure.

What if they fuse anyway?

If you see one hybrid instead of two subjects, targeted intervention is faster than repeating the same prompt.

• Increase the contrast and swap the seed. Give the two subjects stronger, opposite attributes and generate again with a different seed. Often more difference is all it takes.
• Build in layers. Generate one subject first in the photo generator and add the second afterwards, instead of forcing both at once.
• Separate them with inpainting. In the photo editor, select the fused area and repaint one of the two subjects separately, in the same style and lighting.

Because you pay per render and per edit, a targeted fix costs you only that one extra step, instead of running ten full new images hoping the two stay apart this time.

Frequently asked questions

What is concept bleeding?

It's the failure where a text-to-image model lets separate concepts from your prompt overlap or merge, such as two subjects fusing into one hybrid. The term comes from research on multi-object generation, which traces the cause to the text encoder compressing your prompt into a fixed number of tokens.

Why do two animals or two people fuse in particular?

Because lookalike subjects sit close together in meaning for the model. Two animals or two faces share many features, so the model easily combines them in the same spot instead of separating them.

Does adding "not fused" or "two separate" help?

Usually less than you'd hope. Position and contrast work better: give each subject its own attribute and its own place in the frame. That separates the two more strongly than an instruction not to mix them.

How do I stop two faces from merging?

Put the people in separate positions ("left" and "right") and give them clearly different features, like hairstyle, clothing, or age. If it keeps going wrong, generate them separately and combine.

Fusing is annoying but very steerable: make your two subjects different, give them their own place, and correct with intent instead of regenerating blindly. Create an account and test it on your next image with two subjects.