Google charges about six cents per try on image, Kling Kolors about seven, and the free models carry a non commercial license. See prices, quotas, photo rules and a hands on test so you can pick the right virtual try on API for your fashion store or app.
A shopper uploads one selfie, taps an olive trench coat and three seconds later sees themselves wearing it. Behind that tap sits a virtual try on API, and the one you pick sets your cost per image, your photo quality and how much engineering you sign up for. Three options come up in almost every search: Google's try on model on Vertex AI, Kling's Kolors try on, and the free open source models. This article puts them side by side with real prices, limits and license terms, then shows a quick way to test your own garment photos before you write a single line of integration code.
💡 Short answer: Google wins on documentation and enterprise controls. Kling wins on simple pricing and the jump from a try on photo to a video. The free route only wins if your project is non commercial or you already own the GPUs.
A try on API takes a photo of a person and a photo of a garment, then returns a new image of that same person wearing that garment. Face, pose and background stay put. Only the clothing changes. That sounds like a small trick, but it replaces a photoshoot per outfit, per size and per color with a single request that costs cents.
Every provider in this comparison asks for the same pair of images, and the quality of both decides the result more than the model does:
Practical rule: if a human editor could not tell what the garment looks like from your photo, the API cannot either.
Online fashion has a fit problem. Shoppers cannot see how a jacket drapes on a body like theirs, so they order two sizes, keep one and send the other back. Every return means shipping, inspection, refolding and sometimes a write off. A try on feature gives the shopper a believable preview, and a believable preview is cheaper than a returned parcel.
For small brands the angle is different. Shooting each product on three or four models is expensive, so try on turns one model photo into a whole catalog. That is why try on APIs keep showing up in e-commerce roadmaps, and why the choice between providers deserves more than a quick glance at a price list.
Google's offering is a model called virtual-try-on-001, served through the Vertex AI REST API next to the Imagen family. It is the most formal option of the three: versioned, documented, quota based and tied to Google Cloud billing. If your company already runs on Google Cloud, procurement and access control are mostly solved before you write any code.
According to Google's documentation, the contract looks like this:
The output matching your input resolution is a quiet advantage. A 2000 pixel studio photo comes back as a 2000 pixel studio photo, so you may not need a separate upscaling step. Asking for several samples in one request also lets you show a shopper a few variations, or let your own team pick the best one before publishing.
Google bills the model under Imagen pricing. Third party price trackers list about $0.06 per image, and the hosted reseller Pixazo lists $0.063, so budget a little over six cents per output and confirm the number on Google's pricing page before you forecast.
The documented quota is 50 regional online prediction requests per base model per minute. That is plenty for a store and tight for a social app with sudden spikes, so plan a queue and a retry policy from day one.
💡 Check the end date: the model page lists General Availability from January 20, 2026 and a discontinuation date of January 20, 2027. Read the current page before you build, and keep the provider behind a small adapter so that a successor model is a one file change.
Kling's try on comes from the Kolors line and is available through Kling's own platform and through resellers such as fal. The version most developers test is Kolors Virtual Try On v1.5, which uses diffusion based inpainting and works best with front facing model photos and clean garment shots.
On fal the endpoint is fal-ai/kling/v1-5/kolors-virtual-try-on. You send two image URLs and get one image back:
{
"human_image_url": "https://example.com/model.jpg",
"garment_image_url": "https://example.com/olive-trench.png"
}
An optional sync_mode flag returns the image directly in the response. Without it the job is queued and you collect the result later. The example output on the listing is a single PNG at 768 by 1024 pixels, which is fine for a product page thumbnail and small for a hero banner.
That last point matters for brands. If your hoodie carries a small embroidered logo, test it before you promise shoppers an exact preview.
Kling's platform lets you feed a try on result into its image to video model, which turns a still outfit preview into a short clip. You can reproduce that flow on PicassoIA by taking any finished try on image into Kling v3 Video or Kling v2.6 and describing the motion: a slow turn, a walk toward a window, fabric moving in a light breeze.
Short clips like that suit product pages and social ads far better than another static shot, and they cost a fraction of a real shoot.
"Free" has two meanings in this market: free to download and free to run. The open source models are the first and rarely the second.
IDM-VTON is the research implementation behind the paper "Improving Diffusion Models for Authentic Virtual Try-on in the Wild". It ships with a local Gradio demo, inference scripts for the VITON-HD and DressCode datasets, and a public Hugging Face demo for browser testing.
The detail that trips teams up is the license: CC BY-NC-SA 4.0, which prohibits commercial use. That is fine for research, a class project or a personal prototype. It is not fine for a paid store feature. The Kolors family is also published on GitHub and Hugging Face, so read the license of the exact checkpoint you plan to ship.
💡 A paid API at roughly seven cents per image usually stays cheaper than a self hosted pipeline until a dedicated GPU is busy most of the day. Run that math with your own volumes before you commit.
The table below only lists numbers that appear in the providers' own documentation or in public price listings. Where a figure is not published, it says so.
| Factor | Google virtual-try-on-001 | Kling Kolors v1.5 | IDM-VTON |
|---|---|---|---|
| Cost per image | About $0.06 | $0.07 on fal | $0 license, you pay for GPU time |
| Commercial use | Yes, as a paid cloud service | Yes, as a paid API | No, CC BY-NC-SA 4.0 |
| Inputs | Person image plus product photos | Person image plus one garment | Person image plus one garment |
| Output | Up to 4 samples, input resolution | One PNG, example 768 by 1024 | Depends on your setup |
| Speed | Not published in the docs | 10 to 13 seconds per Kling | Depends on your GPU |
| Limits | 50 requests per minute per region, 10 MB per image | Not stated on the fal listing | Your hardware |
| Lifespan | Discontinuation date January 20, 2027 | Not stated | You control it |
| Video follow up | Separate model needed | Kling image to video | Separate model needed |
Before you write integration code, test your garment photos where iteration is cheap. PicassoIA hosts P Image Try On, a clothing swap model that keeps the original face, pose and body shape. It accepts up to six garment images in one run (up to eleven are supported) and returns the result at your original resolution.
💡 At the time of writing, PicassoIA's developer API follows a Replicate style predictions flow but exposes four models, and P Image Try On is not one of them. Use the web interface for testing, and Google or Kling when you need a production API call.
The model page lists these inputs:
Zoom in on seams, hems, collars and any print. Run the same pair again with a different seed if a sleeve looks off. The example runs on the model page mostly finished in under 15 seconds, but a few took several minutes, so any production design should include a timeout and a retry.
For one off outfit changes where you only have a single reference, prompt based editors such as Image Editor Pro or Nano Banana Pro can edit a photo from a text instruction. Dedicated try on models take garment images as direct input, so they are the safer starting point for a catalog.
Most disappointing outputs trace back to the inputs, not the model. Two habits cause the majority of the damage.
A wrinkled hoodie on a messy bed under a harsh lamp gives the model uneven shadows and a distorted shape to copy. Provider documentation says the same thing in different words: single item, clear details, simple background, simple pose. Spend ten minutes on lighting and steaming before you spend money on generations.
A shopper's selfie is personal data, and a photo of a real model is a person who agreed to a specific use. Get permission, say how long you keep uploads, delete them when the session ends if you can, and label generated previews as previews. Check the rules that apply in your market, because they differ. Google's output supports watermarking and Content Credentials, which makes labeling easier.
You now have the numbers: about six cents per image on Google, seven cents on Kling through fal, and a free license that stops at the commercial line. The fastest way to find out which result looks right for your catalog is to run your own photos through them.
Open P Image Try On on Picasso IA, upload one model photo and three garments, and compare the output with what your shoppers expect to see. Then take the best result into Kling v3 Video and watch the outfit move. Experiment with different poses, lighting and garment photos, and keep the setups that work. Your first test costs a few minutes, and it will tell you more than any comparison table.
