Why AI Models Keep Stealing Faces or Turning Bunny Suits Into Dresses

Trying to get AI image editing models like Minimax and Qwen to swap outfits is harder than it looks. You want the clothes from one character on another without transferring their face or body, but these models constantly bleed identity into the output.

This post explores why these AIs struggle to separate clothing from identity. We’ll look at how Minimax and Qwen handle prompts, the hallucinations they produce like adding socks when there aren’t any, and the specific weirdness of bunny suits turning into dresses.

The Prompt Struggle: Identity vs. Clothing

When working with models like Minimax H3 or Qwen Image Edit, the biggest hurdle is that they are trained to recognize whole characters, not just objects. If you ask them to transfer an outfit, they tend to copy the entire person—face, hair, and body shape.

The solution lies in how aggressively we define the boundaries of the prompt. For Minimax, the trick is using explicit phrasing like «Extract ONLY the outfit» while strictly listing what must be preserved on the target character.

With Qwen, a similar strategy works by defining separate variables for the target and the clothing reference. Whether you use long prompts or short ones, the key is to explicitly forbid merging identities. Even simpler phrasing often outperforms complex instructions because it reduces the cognitive load on the model.

The Bunny Girl Hallucination

One of the most frustrating issues arises with specialized outfits like bunny girl suits. When transferring these designs, models sometimes morph them into generic dresses or modify the cut entirely.

This isn’t necessarily a copyright issue, but rather an interpretation failure. The AI recognizes «bunny girl» as a concept and tries to reconstruct its own version of that archetype, ignoring your specific request for an exact transfer.

To combat this, you must treat the outfit as a closed inventory. By explicitly stating that missing items like shoes or socks must not be added, you force the model to respect your reference image.

Ultimately, fighting these models requires patience and precise framing. You have to constantly remind them what is allowed and what is forbidden.