Nano Banana Pro and the AI models behind virtual try-on apps

Ask a shopper what makes one virtual try-on tool better than another and most will point at the result: does it still look like me, did the couch behind me change color, does the shirt look worn or pasted. Almost none of that comes down to the app's interface. It comes down to which AI model is generating the image, a detail that rarely makes it into any product page.

What is Nano Banana Pro?

Nano Banana Pro is not Google's official product name. It is the nickname that stuck after the community started using it for Gemini 3 Pro Image, Google's flagship image generation and editing model, available through the Gemini API and Vertex AI. Give it a photo and a text instruction and it will edit that photo, adding, removing, or changing something while trying to leave the rest alone.

That last part is what makes it usable for virtual try-on, even though clothing was never its one job. A shopper's photo plus a garment photo plus an instruction to swap the outfit is, structurally, the same kind of edit as any other Nano Banana Pro request. Perplexity's own Virtual Try On feature runs on this same model family, a detail Perplexity has stated directly and one already covered in Corlen's own piece on virtual try-on inside AI shopping search.

How is a general image model different from a dedicated try-on model?

A dedicated virtual try-on model is built around one narrow task from the start. The best known example, IDM-VTON, uses a two-stream diffusion pipeline: one stream reads the garment, one reads the person, and a warping step maps the clothing onto the body's specific pose before the final image gets rendered. Every part of that system exists because clothing, and only clothing, was the design brief.

• Dedicated models (IDM-VTON, FASHN, CatVTON, and similar): trained specifically on person-and-garment pairs, with explicit pose and warping steps built into the pipeline. Strong at holding a literal, controllable match to a reference pose.
• General image models (Nano Banana Pro, and the broader Gemini 3 image family): trained on wide-ranging photo editing and reasoning, then applied to try-on through prompting rather than task-specific training. Strong at understanding a full scene at once, not just the clothing region.
• Neither is a strict upgrade over the other. A dedicated model can be more literal about pose. A general model tends to keep the rest of the frame, the face, the room behind it, steadier, because it was never taught to treat those as separate from the task.

Which approach actually produces better try-on results?

There is a real, named answer here rather than a guess. Ionio.ai published a 2026 benchmark testing five virtual try-on models, Kling, FASHN, CatVTON, Qwen, and Nano Banana Pro, across 160 real outfit combinations, scoring pose preservation, fabric realism, identity retention, and production reliability. The finding was not a single winner. Nano Banana Pro came out ahead on overall visual quality and was the pick recommended for creative, photo-realistic use, while the benchmark also noted it lacks the explicit pose and inpainting controls the dedicated models were built around, so its fine-grained controllability trails purpose-built systems on that specific axis.

That tradeoff maps directly onto what goes wrong in a bad try-on result. A pose-locked dedicated model can nail an exact stance but drift on the face or the background around it, since those were never its trained focus. A general model tends to hold the whole photo together, at some cost to exact pose fidelity in edge cases. Which failure mode matters more depends on the job the tool is actually for.

Which models are shopping tools actually built on right now?

This is not a hidden detail once you know where to look. Between the platform launches of the past year and what individual vendors have said publicly, a rough map of who runs on what has become visible.

1. Perplexity's Virtual Try On runs on the Nano Banana image model family, stated directly by Perplexity.
2. Corlen's kiosk, Shopify app, and developer API all run on Gemini 3 Pro Image, the same Nano Banana Pro family model, through Vertex AI.
3. A cluster of smaller, apparel-specific vendors build on dedicated try-on models such as FASHN, CatVTON, or their own fine-tuned versions of similar architectures.
4. Google's own Try-On tool inside Search, Shopping, and Images has not published which specific model it runs on, though it sits inside the same Shopping Graph infrastructure Google has built its broader AI shopping push around.

Which model does Corlen use, and why?

Corlen generates every try-on with Gemini 3 Pro Image, run through Google's Vertex AI, the same Nano Banana Pro family model discussed above. That was not the first thing tried. Earlier development work went through several dedicated try-on models before settling here, and the general image model kept a shopper's face, fabric texture, and the room around them steadier across the kind of photo a real shopper takes: a phone camera, imperfect lighting, a normal room, not a studio backdrop.

That same tradeoff is why how virtual try-on works and is virtual try-on accurate both keep landing on the same two questions worth asking about any tool: does the result still look like the actual person, and does the fabric look worn instead of stuck on top of the photo. The model behind the interface is the real answer to both.

See the difference on your own photo: try Corlen for free, no account needed for a first result. Teams building their own shopping experience can reach the same engine directly through the developer API.