Japanese AI Girlfriend: We Tested 4 Apps With the Same Character

We built the same character four times. A 22-year-old woman living in Tokyo, reserved rather than forward, ordinary rather than glamorous. Then we asked each one the same three questions.

One app let us select Japanese as a distinct option rather than folding it into Asian. One produced a street scene detailed enough to be recognisable. One never let us finish the character at all without entering card details. And the app with actual Japanese characters in its library gave the most touristy answer of the four.

This is what happened, in the order it happened.

How We Tested

We created new accounts on four services and used each platform’s own character builder, matching the settings as closely as each interface allowed. Where a setting did not exist on a given app, we noted the absence rather than substituting something else — the differences between the builders are part of the result.

The character brief was identical throughout:

• Name: Aoi — Age: 22
• Ethnicity: Japanese where offered, Asian where not
• Appearance: ordinary, not styled like a model
• Personality: shy, a little romantic, occasionally teasing; reserved rather than forward
• Context: lives in Tokyo

Then the same three questions, in the same order, with no other prompting:

• I just moved to Tokyo, so I don’t really know anyone yet.
• What do you usually do when you want to relax?
• What part of Tokyo do you like most?

We did not buy credits, submit card details, or work around any paywall. Where a payment requirement appeared, we stopped there and recorded it. We are not scoring how “Japanese” a generated face looks — that is not a judgement worth making, and not one a reader can verify. We are reporting what each builder offered, whether the settings we entered were respected, and what came back in conversation.

Quick Results

What Each App Actually Let Us Choose

Before any character exists, the builder decides what is possible. This is where the four services diverge most sharply, and it is visible in about thirty seconds.

Only one app names Japan

OurDream’s ethnicity screen offers twelve options, and Japanese sits as its own choice alongside Asian — the only country named anywhere in the list. There is no Chinese, no Korean. Just Asian, and then Japanese.

The two thumbnails are visibly different from each other, which suggests the distinction is not cosmetic. Selecting Japanese then changed the rest of the builder: the hairstyle screen switched from the previous background to cherry blossoms and kimono, with eight styles including a bun and a bangs option.

Joi AI offers six: White, Asian, Black, Latina, South Asian, Middle Eastern. South Asian is separated out, but the whole of East Asia stays as one. Lovescape offers the same granularity in Japanese-language UI — アジア系, nothing narrower.

Candy AI takes a different route again. Rather than a fine-grained ethnicity selector, it has Japanese-motif characters already in its library, which is how our earlier character came about — the full account of that is in our first-hand write-up of building a Japanese AI companion in Tokyo.

Nobody let us make a hairdresser

The brief called for a hairdresser, and no app could deliver one on a free account.

OurDream has a job list — nurse, teacher, doctor, lawyer, entrepreneur, athlete, plus a substantial adult-industry section — and a Custom option locked behind a subscription. Joi AI’s list runs to seventeen or so (real estate agent, bartender, librarian, flight attendant, university student, spy) with no custom field at all.

Look at what the lists have in common: nurse, teacher, doctor, maid, flight attendant, secretary. These are roles with uniforms. A hairdresser has no costume, and so has no place in a catalogue built around visual archetypes. That is a design decision about what these products are for, and it is more revealing than any feature list.

Candy AI

Candy AI is where this whole line of testing started, because it was the first app where a Japanese character felt like a deliberate option rather than an accident of the ethnicity dropdown.

The conversation is competent. Setting the personality to reserved produced measured behaviour — small gestures, questions asked and then space left for an answer. Then we asked the third question.

Aoi: I really love Shibuya. The energy there is amazing. *shifts closer, lowering her voice slightly* There’s always something fun happening. Plus, the food is to die for! Want me to show you around sometime?

Shibuya is not wrong. It is also the answer a guidebook gives. Earlier in the same conversation, asked what she does to relax, the reply was gaming and TikTok — plausible for a 22-year-old in Tokyo, and equally plausible for a 22-year-old anywhere.

Finishing a custom character displayed a cost of 10 coins, with a face reroll at 1 coin. We did not see a card requirement in the flow we tested.

OurDream

This is the app that named Japan, and the one that behaved most specifically because of it.

The generated image placed the character in a narrow Tokyo side street: overhead cables, a red cross pharmacy sign, evening light, people walking away from the camera. As a piece of urban reconstruction it is convincing.

The outfit is not what we asked for. We specified casual everyday clothes; the job preset put her in a medical coat and the preset won. That is worth knowing before you assume a character builder will honour a text instruction over a dropdown.

One more thing happened during creation. Our description used the phrase ordinary young woman, and the generator refused it outright:

This prompt was blocked — it may produce content against our guidelines. Content involving minors isn’t allowed: “young”

The coins were refunded automatically. The word was flagged despite an explicit age of 22 elsewhere in the same character, and the system told us exactly which word caused it — a better failure mode than a vague rejection.

Then the conversation. Asked about Tokyo:

Aoi: Honestly? I love the quiet corners. The back alleys in Shimokitazawa, the old bookshops in Jimbocho, places where the city feels smaller. Shibuya’s fun, but it can swallow you whole if you’re not careful.

Jimbocho is Tokyo’s second-hand book district, and the reply knows that. It also declines Shibuya explicitly, which is the opposite of the guidebook answer.

Image generation ran at 5 coins, with a further generation priced at 10, and the custom prompt field carried a Premium badge. The free balance is not published in a form we could verify, so we are not quoting one.

Joi AI

Joi AI could only offer Asian. It then gave an answer nearly identical to OurDream’s.

Aoi: mm… i always liked shimokitazawa… it’s kind of quiet and full of little bookshops and cafes…

Same neighbourhood, same reasoning, from a builder with a coarser ethnicity setting. Whatever produced that answer, it was not the dropdown.

The writing style is completely different from the other two. No prose description, no asterisked gestures — instead lowercase, ellipses, a stray ^^. It reads like a messaging app rather than a novel, and it is the most convincing rendering of shy among the four.

Two things did not hold. The profile reads “Aoi, 20” despite an age of 22 being entered. And the character introduced herself as having just moved to Tokyo for school, then a message later said Tokyo was where she was from — the university-student job preset colliding with the scenario. Activating the custom character displayed a cost of 500 Neurons.

Lovescape — We Never Finished

Lovescape has the most elaborate builder of the four: eight stages, a heavy emphasis on physical detail, four relationship options of which two are step-family. Its Japanese-language interface translates the product as AIセフレ, which is considerably more direct than the “AI companion” wording the others use.

We completed every stage. At the end, instead of a character, we got a checkout.

The page sat on checkout.upgate.com, a different domain from the site itself, and asked for a card number, expiry and CVV. It showed Premium, $19.99 / 0 month and Total $19.99, with a smaller line underneath reading After 3 days, renews at $19.99/30 days.

Read those two together. The largest figure on the screen is $19.99 presented as a total; the three-day free period is in the small text. Both statements are on the page. Only one of them is what your eye lands on.

We stopped there, so we have no Lovescape character to show and no basis for judging its output. That is the result rather than a gap in it: among four apps, one required a payment method before a character could exist at all.

The Finding We Did Not Expect

Set the three answers side by side.

The app with the coarsest ethnicity setting gave a locally specific answer. The app with Japanese characters gave the answer on the postcard.

Two apps independently landed on the same neighbourhood, described the same way, having been given no hint beyond the word Tokyo. That is not the character builder at work — it is the language model underneath, and both of them associate quiet-Tokyo-with-bookshops with Shimokitazawa. Whether that reflects genuine local knowledge or a well-worn path in English-language training data, we cannot say from three conversations.

What we can say is the practical part. What the builder lets you select does not predict what the conversation knows. A dropdown with Japanese in it is a real difference in the interface. It is not a promise about the character you will end up talking to.

What Did It Actually Cost to Create Her?

Free registration is not the same as a finished character. Three of the four apps put a price on the final step; the fourth put a card form there.

We are not quoting starting balances. Each app grants some amount of credit at signup, and we could not verify the current figures well enough to print them — a wrong number here would be worse than none. Check the balance shown in your own account after registering.

For how the four services price subscriptions and credits once you are past this point, our comparison of what AI companion apps actually cost works through each pricing structure with verified figures.

Which One Would We Use?

The honest summary: OurDream produced the most specific result, and it is also the app that made us pay attention to the word “young” and overrode our outfit. No app in this test did everything we asked. Three let us find that out for free.

What We Could Not Verify

• Starting free credit on any of the four apps. Each grants something; we could not confirm current amounts.
• Whether Lovescape’s checkout appears for every user, in every country, at every moment. We saw it once, in one flow.
• Whether the 500-Neuron and 10-coin figures are fixed, or vary by region and promotion.
• How any of these characters behave over weeks. This was one session each.
• Any judgement about which generated face looked most Japanese. We are not making that call.

Costs, credit grants and paywalls in this category change frequently. Treat every figure here as what was displayed on the day we tested, and confirm the current terms on the service itself.

Related Guides

• Building a Japanese AI companion in Tokyo — the first-hand account
• What AI companion apps actually cost, compared
• Candy AI review — features, pricing and privacy
• What is an AI girlfriend? How these apps work