How to identify an anime character from a picture

You have a picture of an anime or game character and no idea who it is. Maybe a repost lost its credit. No single tool answers this reliably, so work up a ladder: start with the cheapest and most certain methods, and move on only when one fails.

The rungs are clues in the picture and the page it came from, general reverse image search, trace.moe for episode screenshots, our Anime Character Tagger, and then confirming the name. This guide is about drawn and animated characters, and none of these methods should be used to identify real people.

What to try first

1. Look for the source yourself

This costs nothing, and it is the only method whose answer comes from whoever posted the picture.

• The page it came from. Read the title, caption, hashtags and replies. People often name the character and the series in the comments even when the post does not.
• Text in the picture. A signature, username or site name is a search term. It usually leads to the artist's own page, where the original post may say who the character is.
• The file name. A character, artist or title sometimes survives in the name. Random characters tell you nothing.
• Metadata inside the file. Image files can carry title, description and author fields, and generated images often carry their prompt. Open the file in our metadata viewer to see. A prompt that names a character is a useful clue, but it is only what someone typed (see how to read AI image prompts). Screenshots and re-uploads often have none, so finding nothing is normal.

2. Run a reverse image search

A reverse image search takes a picture instead of a text query and looks for the same image, or related ones, on the web. You are really asking for pages that contain the picture, which often hold the title, the artist or a caption. Each service has its own index, so a miss on one does not rule out another.

• Google. Google's search-by-image help page says you can use Search by image on google.com to upload a file, drag one into the search box or paste an image link. In Chrome you can right-click an image and choose Search with Google Lens. On Android, touch and hold an image in the Google or Chrome app. Results can include websites that have the image or a similar one.
• Bing Visual Search. Bing Visual Search accepts a dragged image, an upload, a pasted image or link, or a photo from your camera.
• TinEye. You can search TinEye by upload, paste, drag and drop, image URL or browser extension. Its help center explains that it matches a digital fingerprint of your picture against an index built by crawling the web. It finds copies that were cropped, edited or resized, but not typically similar images, meaning a different picture with the same subject.

That last point decides which to use. TinEye is good at tracing one specific picture back to an earlier or larger copy, which is often where the credit lives. It will not find other art of the same character. Google and Bing also list similar images, which may include other pictures of the character with a name attached.

Reverse search only knows what has been indexed, so a brand-new post may return nothing. It can also mislead, because a similar picture of a different character is still a similar picture. Treat a name from one page as a lead and confirm it (step 5).

3. Use trace.moe for episode screenshots

trace.moe searches by screenshot. Its API documentation lists what a result contains: the matching show (identified by its AniList entry), the file name, the episode, the start and end time of the scene, a similarity score and a preview clip. None of the documented fields is a character name. You get the show, the episode and the moment, and from there you can look up the cast.

The trace.moe FAQ says the algorithm is designed for an almost exact match, not a similar one, and that it compares the color layout of the image. In practice:

• Works: a full, uncropped frame (crop to 16:9 or 4:3 and trim extra borders) of at least 320 by 180 pixels, with reasonably sized subtitles. A GIF frame is fine, and the FAQ says a drawing can work if it closely matches the original scene.
• Fails: cropped, flipped, mirrored or rotated images, heavy tints and filters, large text over the picture, dark or mostly plain images, and real photos of anime. Non-Japanese animation is not supported.
• Coverage: most Japanese anime since 2000, some 1990s series and little earlier. New episodes are added about 24 hours after broadcast.

So it is the wrong tool for ordinary fan art, which has no matching frame. Check the score too. The API documentation says results with a similarity below 90% are most likely incorrect.

4. Try the Anime Character Tagger

When there is no source and no episode to find, such as fan art you saved years ago, our character tagger may name the character from the drawing alone.

How it works

It runs the WD14 ConvNeXt V2 model, which learned from Danbooru, an image site where volunteers tag every picture, including who appears in it. For each picture the model scores 2,132 character tags. A character made the list only if it had at least 600 tagged pictures on the site, and the training data ends at the close of 2022. So it knows popular characters with lots of fan art, and nobody who became popular later.

How to use it

1. Check the list first. The search box on the page covers every name the model knows, so you can see whether your character is there before you download the model.
2. Drop, choose or paste the picture. The model downloads when you add the first one (about 370 MiB), and your browser keeps it for later visits.
3. Read the Characters section. By default it names a character only at 85% or higher. The "How sure must it be?" slider runs from 30% to 99%.
4. If the character is small in a larger scene, crop to the character. Every picture is padded to a square and shrunk to 448 by 448 pixels before the model sees it.

What it cannot tell you

The tagger names characters. It does not output an artist, an episode or a series, and it cannot say whether a picture is official art. Some tags carry the series in brackets, as in ganyu_(genshin_impact), when a name could mean more than one character, but many have no bracket, so look the series up. The model card publishes one accuracy figure for all tags together (an F1 score of about 0.69) and none for characters alone, so we will not quote an accuracy number for character recognition. Our anime tagger shows the same results with the descriptive tags first.

How to read a low-confidence result

The page tells you which situation you are in.

• A named character at 85% or more. A strong lead, especially for a well-known character in a usual outfit. Still confirm it.
• "No confident match. Closest guesses:" Nothing reached your setting, so the page lists up to three names that scored at least 15%. Treat them as search terms: they are often a lookalike or a character from the same series.
• "No known character found." Nothing among the 2,132 scored even 15%, so the character is probably not on the list.
• Two or more names. Several characters, or two lookalikes the model cannot separate. Crop each character and run them separately.

If you lower the slider, a name that clears the new setting looks like any other result, so a 35% name is displayed the same way as a 95% name apart from the percentage. Read the number. Scores are separate estimates for each tag and do not add up to 100%, and 60% does not mean exactly a 60% chance of being right. Our confidence threshold guide explains why.

Why recognition fails, and what to do next

The list is a snapshot of its training data. Searching it for frieren, the lead of an anime that began in 2023, finds nothing.

When every method fails, describe the picture and ask people. Note where you found it, the hair and eye color, outfit, accessories, any text and your best guess. The tagger's general tags give you words to use. A fan community for the genre you suspect can often name a character from a description like that.

5. Confirm the name

Never rely on a name from one source. Two independent clues that agree, such as a caption on the original post and a matching design on a wiki, are far better.

• Open a fan wiki for the series. The character's page lists hair and eye color, usual outfit and alternate forms. It also helps you catch a lookalike from the same series.
• Check the series' official site. Its character pages show the official art and name spelling.
• Compare design details, not the overall impression. If the tagger says hatsune_miku, for example, look for long teal twin tails. Fan artists do recolor hair and change outfits, so one mismatch is not proof of a wrong answer, but several are.
• Mind the name format. Tags are lowercase with underscores, and Japanese names usually put the family name first, so hakurei_reimu is the character many English sources write as Reimu Hakurei. See Danbooru tags explained.

Credit the artist

The character belongs to a studio or publisher, and the picture to whoever drew it. If you found the original post in step 1 or 2, link to it, name the artist and leave any signature in place. If you cannot find the artist, say the source is unknown instead of guessing. Credit is good practice, but it is not permission to repost someone's work, so check what the artist allows.