New volunteer project: We're using AI to create mnemonic images for every radical, kanji, vocabulary. Come join us!

WaniKani
21

Hey gills, sorry for that! I got the server back online and you should be able to submit this awesome creation now. :slight_smile: I also confirmed that the Tampermonkey JS script is working too, you can see it in action e.g. on this page: kanji/起

image
image1815×1047 63 KB

I also published a new version of the Tampermonkey script that makes it work properly on the new review pages (I only updated the list of matching URLs, but I tested it and it works fine).

Please let me know if you need any help with anything at all, you can ping me on the server, I am Chiara#9001!

2 Likes
22

Brilliant, thanks so much for all of this!

23

DALL-E 3 is now out and it’s particularly well-suited for generating mnemonic images! I’m going to be doing a new post to recruit some more for this project!

24

I’m building a GPT tool that generates these mnemonics: https://chat.openai.com/g/g-D02cbqhwr-kanjiteller/

It’s got some entertaining results

25

bumping this up in hopes that some people can help with the vocab words :smiley:

1 Like
26

This is fantastic - I just had the same thought today. It would be nice to be able to add your own images to this.
For example - I wanna remember: 池
From ideogram: https://ideogram.ai/g/8bDXO6JMQh6XJ6zEv1RKVQ/3

Edit I guess you could use the reading notes as a means to store “meta data” about the mnemonic. Something like this:

image
image1276×1031 148 KB

Big caveat though - in this case ideogram might cut off hot linking their images,

Edit - I created my own fork - it’s just a code gist:

27

There’s been some discusison before of allowing individual users to customize their own mnemonic images with their own URLs and your idea of putting it in Meaning Notes / Reading Notes is a pretty neat way of doing it so we don’t have to keep track of them in our own database. Might take a look at this later. Thank you for the feature request!

Keep in mind we can also replace our existing images with new ones if the new ones are better! Just join our Discord server and submit your new replacement, and ping me and I can make it the new default. Or if you prefer you can just paste or link the image in Discord and ask me to submit it for you. We’re always happy to get more help with the project. :slight_smile:

28

+1 to chiaracoetzee on joining the Discord server and proposing better images when you have them.

That said, if you are still interested in custom mnemonic images (e.g. maybe you have your own mnemonic that doesn’t match the WK one), I modified the userscript a while back to extract out custom jpg/png images from the Meaning and Reading notes fields as you described. It’s hosted here: raw.githubusercontent.com/aaron-meyers/wanikani-mnemonics/main/userscript/wanikani-mnemonic-images-script.js

I discussed merging this back into the main script in the project Discord but hadn’t heard much interest so I just left it in my repo for now. There’s also support for multiple images from your notes with an image carousel although you must use the mouse wheel or swiping on touch devices to scroll through the carousel (no buttons or keyboard support).

29

Hi everyone, I wanted to reply here about a very similar project I just did to this one. I’m an engineer and sort of ran away with the idea of making an AI-image system for WaniKani on my own.

I did a lot of testing and research around generating images and scripting things to auto-create a ton of the kanji mnemonics. I’d love to collaborate with this group and see if anyone is interested in the work I’ve done so far. I’m happy to share what I did and answer questions. For me, this was an exercise in programming, AI and writing about my engineering process more than anything. I wrote a long breakdown of what I built and why in an article here

Feel free to try out the chrome extension I built here - (Works for levels 1-4 presently).

I realize there is a lot of overlap, and that my approach might not be exactly the same as what this volunteer based project is doing - but I’d love to trade thoughts and ideas with anyone that is interested. Who knows, perhaps there is room for our tools to work together somehow?

6 Likes
30

Cool - you and I are on the same page.
The caveat is that ideogram doesn’t have image urls - here’s an example I made ( only because I got it 中々 wrong again :wink: )
https://ideogram.ai/g/80g2q_9fS--1VAUj83-o2Q/0
I posted my gist in the edited post above.

31

Man, that is super impressive. Automation all the way down. Well done.

1 Like
32

Does anyone have any tips for the best tool and prompts to use to more easily generate images that copy the layout and dimensions of the original kanji?

For instance, for 凍 I would want to generate a flag posted in the ground with exactly this shape 東 and exactly two big drops of ice dripping in the air to the left of it.

33

We’ve been following a volunteering-based approach mainly because I want a human in the loop to be not only vetting all the images before posting them but also doing some prompt engineering to get around the fact that a lot of the text mnemonics on WaniKani are just… not great as image prompts. Some of them deal in abstract concepts that are hard to represent visually for example. But if your generated images do include some of the kanjis or vocab that we haven’t covered yet, it would be a great to do a review pass and consider uploading to our collection those that meet the bar as being good-quality mnemonics. Thanks for working on this. :slight_smile:

34

Hmmm, if it’s only script changes and it’s not disruptive to users who aren’t using the feature I’d be happy to merge your changes into the main script, please ping me on the Discord about it. :slight_smile:

35

For some guidance on generating images that match the layout of the kanji, see what they do with Stable Diffusion and QR codes which is pretty similar:

How to generate a QR code with Stable Diffusion - Stable Diffusion Art

Essentially you can use a base image that is just a large version of the character made in Photoshop etc. as the basis of img2img proceessing.

1 Like
36

I’d better be able to turn these off. I hate this idea and it frankly makes me ill to consider it.

37

This is a volunteer project by users unaffiliated with Tofugu. You won’t see them unless you install the user script.

2 Likes
38

Ah, I apologize. I did not get much sleep last night. Still, as an artist, I take exception to the whole operation anyway.

39

I love and support human artists but none of can afford as individual users to hire commissions to draw 18000+ images to cover all the content. That would cost millions of dollars and take years. We really just want to help everyone learn kanji better.

1 Like
40

Yes, “we can’t afford it otherwise, we just want–” is why the companies who collected these databases from artists didn’t pay the artists in the first place. It’s why cotton plantation owners in the states justified all kinds of practices, it’s why the coca cola company and amazon have gone to such lengths to squash unionization, and it’s why thousands of companies and countries for the past thousands of years of human history have justified all kinds of things. You may choose to justify it however you like. I guess what you do, in the end, is between you, your conscience, and God.