When a client uploads a photo on a salon's booking page powered by Ligax, a lot happens in a fraction of a second. Lig Vision, the model that powers our platform, looks at your image and works out what style you have, what technique was used to create it, how it was installed, and what extras or accessories are part of the look. That information is what connects you to the right stylist, at the right price, for the right amount of time.
Getting that right matters. A misidentified style means the wrong stylist, the wrong quote, or a booking that does not match what you actually want. So when we say our model is accurate, we think you deserve to be able to verify that for yourself rather than simply take our word for it.
That is exactly why we built and published the Ligax Hairstyle Classification Benchmark.
What Is the Benchmark?
A benchmark is a standardised test designed to be reproduced independently. We selected a focused set of 314 photographs from one of the hardest style categories our model works with, had each image labelled by professional hairstylists with active working knowledge of the domain, and then ran that test openly against three of the most capable general-purpose vision models available today.
Every test image, the labels applied by our professional stylists, all the results, and all the code used to run the test is publicly available. Anyone can download it, run it themselves, and check our numbers independently. If you want to verify a specific image, you can upload it directly to our playground at ligaxai.com/playground and see what Lig Vision returns.
One thing worth saying clearly: this benchmark is not a complete picture of what Lig Vision does. The model recognises hairstyles across a wide range of cultural and regional traditions, including West African, East African, Caribbean, South Asian, East Asian, and European styling. It operates across many more style categories than this benchmark covers. We chose one specific domain for this first public test because it is the hardest to get right and therefore the most meaningful place to start. The benchmark is a window into performance, not the full view.
Who Did We Test Against?
We ran the same test against three leading general-purpose models: GPT-5.5 from OpenAI, Claude Opus 4.8 from Anthropic, and Gemini 2.5 Flash from Google. These are amongst the most powerful systems in the world today. None of them was built specifically for hairstyles. They received the same images and the same instructions as Lig Vision.
We included them because any meaningful benchmark requires a credible baseline. The best available alternative to a specialist model is one of the most capable general-purpose systems in the world. If Lig Vision only barely outperformed tools built for entirely different purposes, that would tell you something important about the value of domain-specific training. As it turns out, the gap is substantial.
Why Protective Hairstyles
We focused this first benchmark on Afro-Caribbean protective hairstyles deliberately. These styles are technically amongst the hardest to classify correctly from a photograph, for reasons that are worth understanding.
Many protective styles look very similar at a glance. Box braids, faux locs, marley twists, and passion twists can share the same length, the same colour, and even the same overall silhouette. What separates them is the texture of the extension material, the geometry of how the base is installed, the thickness of each strand. These are details that require genuine domain knowledge to read from an image.
By starting the benchmark here, at the demanding end of the difficulty range, we get results that are maximally informative. A model that performs well on this test can be expected to handle less challenging style categories effectively. The reverse is not guaranteed.
The Dataset
The 314 photographs used in this test were collected from publicly available sources. Every single image was labelled by a professional hairstylist with active, working knowledge of the styles involved. Each label set was then independently reviewed by a second professional. Any disagreements were resolved before the labels were finalised.
Crucially, Lig Vision has never been trained on any of these images. The dataset was assembled after training was complete, and our test fetches the photographs directly from their public source at evaluation time rather than from any stored copy. The test is clean, and can be independently reproduced by anyone.
The Results
Lig Vision outperformed all three comparison models by a substantial margin on both measures reported.
For identifying the specific named style, Lig Vision scored 0.622. Claude Opus 4.8 scored 0.477. Gemini 2.5 Flash scored 0.433. GPT-5.5 scored 0.411.
For identifying the base technique category, Lig Vision scored 0.780. Claude Opus 4.8 scored 0.693. Gemini 2.5 Flash scored 0.667. GPT-5.5 scored 0.572.
Scores run from 0 to 1, where 1 represents getting every image exactly right. On the harder part of the test, identifying the specific named style, Lig Vision outperformed the nearest general-purpose model by roughly 30%.
It is also worth noting that these results reflect a constrained public evaluation on a single style domain, not the full scope of the model's capability. Where the numbers are lower, this often reflects the difficulty of specific styles that are genuinely hard to distinguish even for human experts, or style classes with very few examples in the dataset rather than any weakness in the model's overall performance.
What This Means for Salons and Customers
For a salon, accuracy at the identification step translates directly into bookings that arrive correctly described. The stylist knows what is coming. The right time is blocked. The right materials are prepared. There is no awkward conversation at the appointment about what the client actually wanted.
For a customer, it means the service selected during booking is the service received. The match between the reference image, the service in the catalogue, and the price shown is grounded in a model that has been specifically trained on this domain, tested against expert labels, and publicly benchmarked.
The Benchmark Is Open
We have published everything. The dataset is publicly available, the evaluation code is on GitHub, and every result can be verified independently. If you want to check a specific image yourself, visit the Ligax playground at ligaxai.com/playground, upload a hairstyle photograph and see what Lig Vision returns. The same model that produced the benchmark results powers every booking page.
The full benchmark is available at github.com/Ligax-LTD/Lig_benchmark_test.
We will continue running and updating this benchmark as the model develops and as we expand to additional style categories and regions. Every result will be published in the same format, against the same standard. If you are considering Ligax for your salon, or if you are a customer who has ever wondered whether the photo you uploaded actually got read correctly, this is how we answer that question.
Benchmark repository: github.com/Ligax-LTD/Lig_benchmark_test. Verify predictions at ligaxai.com/playground. Questions: admin@ligaxai.com.