Hacker News Reader

Comment by Brajeshwar | Hacker News Reader

I did something similar for my younger daughter using MagicStory[1] (not mine). I have told her the story quite a few times at her bed time. She has shown it to her teacher and her friends are "jealous." It has her pictures, kinda morphed like a princess and stuff.

Edit: Clarifications.

1. https://magicstory.ai

View on HN

The images looks neat but the stories are a little bit too simple/shallow aren't they?

Another thing. Don't you have a bad feeling uploading images of your kids, conveniently tagged with their age, to a random web service that promises to delete the images without even revealing the makers?

I mean yes, it looks legitimate and you can find a person on Twitter who seems to be the maker of it, but that's not a guarantee for the safety of the uploaded images. I'm not a data paranoid, but careless sharing my kids' images with a random service like that crosses a border.

Yes, they feel like all the stories I asked ChatGPT to write.

They are all very bland, and sound like someone is making a summary of the story rather than reading the story itself.

So I think it's great to get ideas of a plot for a story, but then you should write it yourself to make it more "lively".

I agree with this, but also note that a lot of children's books written by humans (I suppose) are terribly bland. My daughter has some Disney books which has illustrations from the films, and the writing in those is exceptionally bland. Same goes with some wonderfully illustrated, but terribly written, books from the 50s and 60s. I'd actually much prefer reading her GPT stories than those books which we have on the shelf (I guess I should get rid of them actually).

There are also a lot of really nice books for children, with great plots and illustrations. And you can actually read them before buying, if you go to a bookstore.

> loading images of your kids ... to a random web service

I used to worry about that... but then someone pointed out to me that there are pretty much no stories of anyone being hurt by that.

Worst case, those images leak all over the web... And then what? The internet is already filled with pictures of people and children, yours won't stand out. The photos aren't tagged by name, and even if they were, few people care about obama as a 5 year old pirate [1], so why would they care about your child?

Overall, the risk of personal harm from a data leak seems really small - so I put my efforts elsewhere, like making sure my child gets to spend more time with me - that's infinitely more valuable.

[1]: https://img.buzzfeed.com/buzzfeed-static/static/enhanced/ter...

Hey! Maker of https://magicstory.ai here! Super glad to hear your daughter loved it :) we're pushing updates to improve the stories and definitely take privacy seriously - right now all images get deleted once your characters model is trained.

Also planning to have stock characters so you can make stories without uploading images since I get that's definitely a concern for most parents. Would love any other feedback on how to improve the experience!

I guess we all had the same idea because I'm building something very similar too :)

When did you get the idea? ChatGPT is what sparked it for me.

Right before xmas! Gave my little sister a story as a gift :)

Are you planing to offer a non-Google registration?

These look amazing! Did you build MagicStory?

Nope. I think it was a Show HN or something of sort that showed up on my radar.

I did :)

Congrats, looks really amazing! May I ask what your AI stack is? Do you use MidJourney for the images and GPT for the stories or something else? How do you ensure the output stays consistent across images - do you use the seed image and a custom set of prompts for each image, or did you use eg StableDiffusion and did some customized training à la Dreambooth to ensure consistency? Is the workflow fully automated, or are the outputs reviewed and the prompts iterated on manually?

Thanks so much! Happy to share our stack: - GPT3 davinci-003 writes the story - we did a lot of manual experimenting w prompts and used that to fine-tune GPT3 which now generates the image prompts for us - Stable Diffusion + Dreambooth generate the images (we train a model on your hero which gets deleted once your story is finished)

Right now the workflow is mostly automated, but we still manually approve stories before they get sent out to ensure quality. Unfortunately bc of manual time it takes + GPU costs + super high demand we will probably start offering a paid tier. Hope that helps!

Really cool idea, but I would feel uneasy uploading my child's picture to your website. All I have is your promise that you're going to delete them later.

By the way, if I may ask: since story generation seems to be free, what is your business model?

Thanks -- totally understand! In a few weeks we will have some starter characters so that you don't have to upload any images to get a story.

Re: pricing bc of the manual time it takes to approve stories + GPU costs + super large queue we actually just started offering a paid tier to move up the queue and get your story within 24 hours. Definitely still experimenting tho!

Great, thanks very much for taking the time to reply, and good luck with this!

Reply on news.ycombinator.com