Khroma

The problem with colour libraries is that they are somebody else's taste, arranged by popularity. You scroll, you find nothing quite right, you scroll further, and eventually you pick something adequate because you have run out of patience.

The reason browsing fails is that people cannot describe the colours they want and can recognise them instantly. A library asks for the wrong kind of judgement.

Khroma's approach uses recognition directly: choose fifty colours you like, and it learns from that choice rather than asking you to explain it.

Khroma: A Palette Generator That Learns You

• You train it — choose a set of colours and a neural-network-powered algorithm learns which to generate and which to block.
• Training runs in your browser, so the personalisation stays on your machine.
• Four output views: typography, gradient, palette, and applied to your own image.
• Search by hue, tint, value, colour, hex and RGB, rather than only scrolling.
• WCAG accessibility ratings on every pair, alongside hex, RGB and CSS — made by George Hastings.

Can An Algorithm Learn Your Taste?

Khroma states that by choosing a set of colours you train a neural network powered algorithm to generate colours you like and block ones you do not, and that it has learned from thousands of the most popular human-made palettes across the internet.

The two halves matter separately. The base model encodes what works generally — the relationships that make human-made palettes cohere — and your training pass filters that towards what you personally respond to.

That is the right division. Learning colour theory from your fifty picks would be hopeless; learning your preferences on top of a model that already understands relationships is a much smaller problem.

It also fits how taste actually behaves. Nobody can articulate why they like a particular desaturated green and dislike the one beside it, but they recognise both instantly, and a training pass captures exactly that unarticulated preference.

The limitation is worth stating plainly: it learns what you liked in isolation, on a screen, out of context. Colours you like as swatches are not necessarily colours that work as an interface, which is a different question the tool cannot answer for you.

Nobody can articulate why they like a particular desaturated green and dislike the one beside it. A training pass captures exactly that.

Training It In The Browser

The site specifies that the algorithm is trained right in your browser, which is a technical detail with practical consequences.

The first is privacy in the plainest sense — your preferences are yours, sitting on your machine, not a profile held somewhere and used to infer things about you.

The second is that it works without an account for the training itself, which removes the friction that kills most personalised tools before they demonstrate anything.

The third is fragility, and it is the one to plan around. Anything that lives in a browser can vanish with cleared site data, a different device or a private window, so the trained model may not be as durable as it feels.

The practical response is to save the combinations you like into a collection and copy the values out into your own tokens rather than treating the trained model as permanent storage — the same discipline that applies to any account-free tool, including ColorKit.

Searching By Hue, Tint And Value

Khroma's search lets you filter the generator by hue, tint, value, colour, and by hex and RGB values.

Search on top of generation is the feature that makes it a working tool rather than an inspiration toy. Generation answers show me something; search answers I need something like this, which is the question you actually have at 4pm.

Filtering by value specifically is more useful than it sounds. Interface colour problems are usually value problems — two colours of similar lightness produce a design that looks flat and text that is hard to read — and being able to constrain by value directly addresses that.

Searching by hex means you can start from a fixed brand colour and generate around it, which is the actual constraint most professional work operates under. A palette generator that cannot accept an existing colour is unusable in a brand context.

The combination of a personalised generator plus constraint-based search is genuinely uncommon. Most tools do one or the other, and the pair is what makes this worth the setup time.

• Spend the full training pass — a rushed one produces a mediocre model.
• Start from your fixed brand colour by searching hex, then generate around it.
• Filter by value when a design looks flat, since that is usually the cause.
• Check the WCAG rating on every pair you intend to put text on.
• Copy final values into your own tokens; browser state is not storage.

WCAG Ratings And Why They Belong Here

Khroma shows colour names, hex codes, RGB values, CSS code and a WCAG accessibility rating for each pair.

Including the contrast rating alongside the aesthetic output is the single most consequential design decision in the product, and it is easy to miss.

The reason is that contrast is the one colour question with a right answer. Everything else on the page is preference; the ratio between text and background is a measurement, and failing it means people cannot read your interface.

Putting the number next to the swatch means you find out at the moment of choosing rather than after a complaint. A tool that shows you a beautiful pairing and lets you discover its illegibility later has actively made your work worse.

The habit worth forming is simple: any pair you intend to put text on gets checked before it enters the design, not after. That single discipline removes an entire category of accessibility defect from a build.

Where A Personalised Generator Stops

There is a boundary to this approach, and knowing it prevents expecting the wrong things.

It generates combinations, not systems. A design system needs colours with assigned roles — background, surface, text, border, accent, error — and a beautiful pair tells you nothing about which is which.

It cannot account for context. The same colour reads differently at a large fill than in a one-pixel border, and a generator working in swatches has no view of how much of each you will actually use.

It also encodes your taste at one moment. Training on what you liked this month means it will keep producing that, which is fine for consistency and unhelpful when you want to be pushed somewhere new.

And it stops entirely at the point where colour becomes a material or a rendered space — the domain of tools like Decory and Collov AI, where light, surface and finish decide the outcome more than any hex value does.

Building A Palette With Khroma

The way people waste the training investment is rushing it, then blaming the generator for producing colours they do not like.

This sequence spends time where it compounds and puts the measurement step before the decisions.

1. Do the full training pass properly — every rushed pick makes the model worse.
2. Search your fixed brand colour by hex and generate around it rather than from nothing.
3. Assign roles deliberately: one background, one surface, one text, one accent.
4. Check the WCAG rating on every text-on-background pair before committing.
5. View combinations as typography, not just as swatches, since that is how you will use them.
6. Save the values into your own design tokens rather than leaving them in the browser.

Khroma: Tool And Maker