If you spend any time with AI image generators, you have probably hit the same wall. You finally generate a character you love, and the next image gives you a stranger with similar hair. You wanted a series of posts, a short video or a comic, and suddenly the main character looks different in every frame. It is probably the most common frustration in AI creative work.
This time I want to introduce AI Portrait Gallery, a resource built for exactly that problem. According to the official site, it is a collection of reference images that helps characters keep the same face across AI-generated images and videos, split into front-facing faces and styled looks, and released under CC0. The whole site is in Japanese, but it is easy to figure out. For creators who need a stable character face, a free, commercially usable library is a generous offer.
A quick note on sources. This article is based on the text of the official page and the public descriptions of several GitHub projects built around it. The figure of 1,499 images comes from those project descriptions, and I did not count them one by one. I also cannot see the usage examples the author shares in a LINE open chat community, so please read what follows as a summary with commentary.
What is AI Portrait Gallery?
AI Portrait Gallery is a free face reference library published by a Japanese creator known as あぎ, offering AI-generated fictional portraits that help characters keep the same face across different images. The Japanese name on the site is "AIモデル顔データベース", which roughly translates to "AI Model Face Database". The operator is listed as あぎ, with a link to the creator's X account. To be clear, this is not an image generator, but a library of reference images you download.
The site splits its images into two sections. "Face" (フェイス) shows front-facing faces, while "Styling" (スタイリング) also shows clothing, hair texture and overall mood. According to the GitHub project descriptions, the two sections add up to 1,499 images, with 893 faces and 606 styled looks. What caught my attention this time is the sheer scale, since more than a thousand options give you room to find one that really fits.
The files are not small either. According to those descriptions, the originals are PNG files ranging from about 1122×1402 to 1597×2000 pixels, and the thumbnails are 400×500 WebP images. All 1,499 originals add up to roughly 4.5 GB, so plan your disk space if you want everything. In practice, most people only need a handful of favorites, not the whole set.
Why do AI characters change faces, and how does a reference library help?
Character consistency means keeping the same recognizable face for one character across different scenes, poses and lighting. Image generators sample fresh randomness on every run, so a text description like "woman in her thirties with short dark hair" still fits thousands of different faces. That is why, no matter how many adjectives you add, the model picks a new face each time. It is also why many guides on character consistency put the focus on reference images, not on prompts.
Think of it like casting an actor. You would not tell a director to find a gentle woman around thirty, you would hand over an audition photo. AI Portrait Gallery gives you more than a thousand audition photos, and every one of them is a person who does not exist. Pick one, bring that face into every generation, and your character is much less likely to be replaced by a stranger.
For anyone making social posts, comic panels or short videos, this need is very real. Imagine a twelve-image post series: if the main character changes face even once, the whole series falls apart. Ten minutes spent choosing a face up front usually saves far more time in retouching and regenerating later. To me, that is where a library like this pays off the most.
How do you use AI Portrait Gallery?
Using AI Portrait Gallery is simple: pick a face, download the original image, and load it into the character or face reference feature of the AI tool you already use. The official usage tips follow the same idea: choose "Face" when you want to lock a face, and "Styling" when you want clothing, hair texture or mood as a reference. Below I break it into five steps, and since every AI tool labels its buttons differently, follow the interface of the tool you use.
Step 1: Decide between a face and a styled look
After opening the site, choose between "Face" and "Styling". If your goal is a fixed face for your character, go with the Face section, and if you also want outfit and mood as a reference, go with Styling. My suggestion is to start with Face, because it has the fewest variables.
Step 2: Narrow things down by hair length
The site lets you filter by hair length: short, medium, semi-long and long. Scrolling through more than a thousand images makes your eyes glaze over, so filtering by hair first cuts the list down quickly. If your character already has a hairstyle, this step can shrink your choices to a few hundred.
Step 3: Download the original-size image
Once you find a face you like, the official tips recommend downloading the original-size PNG so fine details survive. If you use a thumbnail as your reference, eyes and strands of hair turn blurry, and the AI has fewer clues to work with. As for file size, the project descriptions say the average is about 3 MB per image, so downloading a few is no burden.
Step 4: Load it into your AI tool
Next, drop the image into the image or video tool you normally use. The official site only says you can load it into the character reference or face reference features of various AI tools, without naming a specific one. Quite a few tools have a similar entry point, though the name differs from product to product. Then keep your prompt focused on the scene, clothing and action, and leave the face to the reference image.
Step 5: Stick with the same face
The official site notes that using the same image consistently makes it easier for generated images and videos to keep the same face. So once you pick your lead portrait, do not swap it out during the project. When you need new outfits or an expanded character sheet, look in the Styling section for images with a similar feel and use them as clothing and mood references. Keep in mind that "easier to keep consistent" does not mean "guaranteed", and results still vary from tool to tool.
What is the difference between the Face and Styling sections?
The Face section is for locking in a face, while the Styling section is for referencing clothing, hair texture and overall mood, so the two serve different purposes and cannot replace each other. The table below sums up the differences, with counts and hair-length breakdowns taken from the GitHub project descriptions. You will notice the hair-length proportions are very similar, with short hair the largest group in both. For anyone browsing, that means short-haired characters have the widest choice.
| Item | Face (フェイス) | Styling (スタイリング) |
|---|---|---|
| Number of images | 893 | 606 |
| What you see | Front-facing faces on a simple background (per project descriptions) | Clothing, hair texture and overall mood |
| Recommended use per the site | When you want a consistent face | When you want clothing, hair texture and mood as a reference |
| Short hair | 433 | 294 |
| Medium hair | 229 | 147 |
| Semi-long hair | 154 | 106 |
| Long hair | 77 | 59 |
Long hair is the smallest group in both sections, so if your character has long hair, your options are more limited. In that case you could try starting from a semi-long face and adjusting the hair length through your prompt. Of course, results depend on your tool, so test on a small scale first.
What does the license allow, and can you use it commercially?
AI Portrait Gallery images are released under CC0 1.0, which essentially means the author has placed them in the public domain, so commercial use, editing and reference are all allowed and credit is not required. According to the official site, every image is an AI-generated fictional person who does not exist in real life. The site also links to an explanation of CC0 1.0 for the full terms. For freelancers and anyone making commercial content, this is a refreshingly simple license.
The site also lists a few rules, which I have put into the table below. One of them is a request, not a license condition, and that difference matters. The author asks people not to redistribute the image files themselves as they are, which feels like a personal wish from a creator. My own view is that when someone releases work for free, a request like that deserves to be respected.
| Rule | What the site says | My note |
|---|---|---|
| Commercial use and editing | Free to use commercially, edit and reference | No prior approval needed, but the rules below still apply |
| Credit | Not required, and crediting the source is optional | A link to the official site is the easiest way to say thanks |
| Real people | Fictional AI-generated people; impersonating real people or uses against public order and morals are not allowed | Even a fictional face can end up looking like a real person, so be careful with the results |
| Redistribution | Please do not repost or redistribute the image files as they are (a request) | When sharing, link to the official site instead of bundling images |
| Metadata and warranty | Metadata has been removed, and everything is provided without warranty | You are responsible for how you use the results |
One more thing is easy to mix up. Beyond this site there are other galleries in the same family, such as pose and costume libraries, but according to third-party project descriptions, those galleries do not state an open license on their original sites. So CC0 applies only to this portrait gallery, and you should not assume the same for other galleries just because the names look related.
What should you watch out for before using it?
There are three things to handle with care when using this gallery: adult-oriented content, the risk of impersonating real people, and how you share the images. The first is the content itself. According to the description of one Chinese-localized project, the material consists of adult-oriented portraits including swimwear and lingerie looks, with a notice that minors should not browse it. The text on the official page does not show the same notice, so I suggest you look through the Styling section yourself before opening it in a workplace or school setting.
The second is impersonation. The people in the library are fictional, but an AI-generated face can still happen to resemble a real person. The official site explicitly bans impersonating real people, and I do not think that rule is just a formality. If you run a long-term AI character, I personally think stating clearly in the profile that it is AI-generated will put your audience more at ease.
The third is redistribution. The author asks people not to republish the images as they are, yet GitHub already has quite a few mirrors and derivative projects, and some claim to include every image. CC0 imposes no legal limit, but I think honoring the author's request is the better way to treat a creator. So if you want to recommend it to a friend, just share the official link.
How do the GitHub localized versions and spin-off projects differ from the official site?
These spin-offs are localized versions or extra tools that other people built on top of the official site, and they are not the official website itself. For readers who cannot read Japanese, a localized version is convenient because the interface and hair-length labels are already translated. Whether a spin-off stays in sync with the original or translates accurately, though, is something you have to judge for yourself. The table below lists a few examples I found, and the star counts are the numbers shown when I put this together, so they may change.
| Name | What it does | How it differs from the official site |
|---|---|---|
| Official site: AI Portrait Gallery | A Japanese-language face reference library with images released under CC0 | The original source of the images and the rules |
| Brilliant666 / ai-portrait-gallery-cn | A Chinese and Japanese bilingual reproduction that translates hair-length labels and removes the original community promotion links | Its description says the repository itself contains no images |
| anhao / ai-pose-gallery-zh | A Chinese version that merges the pose, costume and portrait galleries, plus an AI character compositing studio and a ComfyUI plugin, showing about 5.3k stars | Much broader than the portrait library, and the pose and costume images are not CC0 (according to third-party descriptions) |
| atuizz / PoseWear | Pose, costume and portrait references plus in-browser AI virtual try-on, using your own API key | A different kind of app, not just a library |
My own approach is to browse a localized version for a quick look, but to download the images I actually use from the official site. That way the license, the source and the author's latest notes are all in one place. Also, always check that a third-party site is trustworthy before following its links or downloads.
How does it compare with other face libraries, and is it worth using?
The value of a face reference library is that ready-made fictional faces save you the time of finding and tuning a face, and libraries differ mainly in size, license and how closely they connect to a generator. Similar services exist, such as the face gallery from WaveSpeed AI, which lists 284 faces of people who do not exist and says they are free for personal and commercial projects. The biggest difference is that WaveSpeed's page offers a button to generate photos of that person directly, while AI Portrait Gallery is purely a library. The table below compares the two based on what each page says, and I have not used WaveSpeed myself.
| Item | AI Portrait Gallery | WaveSpeed AI Face Gallery |
|---|---|---|
| Number of images | 1,499 (per project descriptions) | 284 (per the page) |
| License | CC0 1.0 | The page says free for personal and commercial projects |
| How you use it | Download, then load into another AI tool | Open the generator with the chosen face directly from the page |
| Language | Japanese (plus third-party Chinese versions) | I saw a Chinese-language version of the page |
| Best for | People who want to pick a face and use it with any tool | People who want to pick a face and generate in one place |
To me, the choice depends on your workflow. If you already have an AI tool you like and want lots of options under a simple license, AI Portrait Gallery's 1,499 images give you more room. If you want to do everything in one place without hopping between sites, an integrated service like WaveSpeed is more convenient.
Is this library worth adding to your toolbox?
AI Portrait Gallery is a free, simply licensed library of faces that works well as a starting point for an AI character, not as something that finishes the character for you. What I like most this time is how small a problem it tackles, only "keep the face consistent", and the CC0 license that lets you use it with peace of mind. The downsides are just as clear: a Japanese-only interface, adult-oriented content in the Styling section, and the author's request about redistribution all deserve some care. If character drift is wearing you down, ten minutes spent browsing and picking a lead portrait should not hurt.
For anyone building a long-running AI character, locking in one face takes less effort than endlessly tweaking prompts. And as free resources keep multiplying, what matters is not grabbing more images but remembering to respect the creator's wishes. I think those two ideas belong together.
What are the most common questions about AI Portrait Gallery?
This FAQ answers the questions people most often ask about the licensing, usage and differences of AI Portrait Gallery before they start using it. The six answers below are based on the official site and third-party project descriptions. If your situation is unusual, treat the latest official notes as the final word.
Can I use AI Portrait Gallery commercially?
Yes. According to the official site, the images are released under CC0, so commercial use, editing and reference are all allowed, and credit is not required. The site does ask people not to redistribute the images as they are, and it prohibits impersonating real people or any use that goes against public order and morals.
Are the people in the images real?
No. The official site states that all of them are AI-generated fictional people who do not exist in real life. Still, an AI-generated face can happen to resemble someone, so stay alert when you use them.
How do the GitHub localized versions differ from the official site?
The localized versions are third-party reproductions or extensions built on the official site. Some translate the hair-length labels, and others add a compositing studio or virtual try-on features. They are not official releases, so judge their update speed and accuracy yourself, and download images from the official site when you can.
Does it work for readers outside Japan?
The official site does not mention any regional restrictions, and the interface is in Japanese. I have not tested connections from different regions. If the page does not load, try switching networks.
Is there a paid version, and is it worth paying for?
The official site describes the images as free to use, and I did not see any paid plan on the page. The author also runs a LINE open chat where usage examples and generation prompts are shared, but I cannot confirm whether any other content is paid. For most people the free library is plenty, and you can decide later whether anything else is worth joining.
Will using the same face guarantee identical results?
Not always. The official site only says that continuing to use the same image makes the face easier to keep consistent. Results vary with your tool, prompts and any other reference images, so test on a small scale before generating in bulk.