storystarling
storystarling on Hacker News
I'm working on StoryStarling (
https://www.storystarling.com). It creates custom hardcover children's books from scratch around any idea, instead of swapping names into pre-set templates.
We mostly focus on things you cannot find in bookstores: bilingual stories with parallel text for multilingual homes, and non-fiction for kids who have hyper-specific obsessions (like tunnel boring machines or wastewater systems).
Yes, the illustrations are AI-generated. We generate the book, lay it out for print, and print/ship the physical copy.
Fair point though - "we illustrate" could be clearer.
https://www.storystarling.com - create a non-fiction children's book explaining your super-niche-geek topic to your kid. Pick any topic, your kid becomes the little explorer, we illustrate and print it. Requires registration, but then lets you read the whole book before paying.
I'm working on StoryStarling (
https://www.storystarling.com). You describe an idea for a children's book, optionally upload photos of your kid or pet to put inside, pick a style, preview the result, and order a printed hardcover. Bilingual if you want.
StoryStarling. You describe a story idea and it generates a fully illustrated children's book, then we print and ship it.
Not templates with names swapped in. Every story and illustration is made from scratch. You can go from "dinosaurs soccer" or write out a whole storyline. Pick an art style, optionally upload reference photos of your kid, and it builds a 28 page book in a few minutes.
Bilingual in 38 languages. We handle RTL (Arabic, Hebrew), CJK, and less common languages like Estonian, Maltese, Irish where there's not much available for kids.
Tech side for the curious: LangGraph orchestrates the pipeline, Celery workers do image generation and text rendering in parallel, and LLMs critique the illustrations for consistency mistakes and can trigger regenerations automatically.
Printed in Germany, booklet around 20 EUR, hardcover around 40 EUR.
https://storystarling.com
Appreciate the honesty. On bilingual - that's actually our strongest use case already. A surprising number of our orders are from multilingual families who simply can't find children's books in their language combination. That's a gap no publisher will fill because the print runs are too small.
This matches our experience. For a 28-page children's book there's no context collapse - the entire story fits comfortably in context. The format actually plays to LLMs' strengths: short, structured, clear emotional arc. It's a very different problem than generating 500 pages.
Fair perspective. But the parent isn't passive here — they're the creative director. They decide what the story is about, who the hero is, what happens. The AI does a lot of the writing, yes, but the parent is the editor: they review every page, rewrite lines, regenerate illustrations they don't like. It's closer to working with a ghostwriter than pushing a button.
Most AI content feels empty because it's made for nobody in particular. A StoryStarling book is the opposite - a parent shaping a story around their specific child's world. That's a real story. They just had help telling it.
Honestly, for children's books specifically - yes. A children's book is 28 pages, simple language, short sentences, clear emotional arc. That's a very different challenge than writing a novel. We've put a lot of work into the prompting and story structure, and the results are genuinely good for this format. Parents can also edit every page of text before approving for print if they want to tweak anything.
The soul comes less from the prose and more from the fact that this story exists for this child. A book about their specific
fear, their favorite thing, their family situation — that's what makes a kid ask to read it again at bedtime.
That's a fair point about friction, but we're intentionally focused on physical books - the whole idea is to get away from screens. Most parents we talk to have the opposite problem: too much digital content, not enough tangible things. A printed book that lives on the shelf, gets read at bedtime, gets dog-eared and carried around - that's the product. You're ordering it for a birthday, a holiday, a first day of school. It's not impulse content, it's a keepsake.
yes - we do combine those both and you can upload photos to get your kids into the book.
Thanks for the blunt feedback - genuinely appreciate it.
You're right that children's books can be excellent, and for generic topics a well-reviewed book from a skilled author and illustrator will beat what we generate. No argument there.
Where we see real value is in the gaps the publishing industry doesn't serve. Bilingual families who can't find books in Maltese/English or Estonian/German. A child with an insulin pump who wants to see a superhero like them. A kid processing their parents' divorce. A child with two dads, or being adopted, or starting at a new school in a country where they don't speak the language yet. No publisher will print a run of one for these families - but these are exactly the stories that matter most to them.
On the UX points - you're right on both. We should localize the showcase to your language, and the signup wall before trying is too much friction. Working on both.
Similar experience. I posted a Show HN two days ago for a children's book generator - type a story idea, get a fully illustrated printed book shipped to you. Offered a free printed book including shipping to the HN community via voucher code. Got 7 points, 2 comments, and zero voucher redemptions. Nobody even ordered the free book.
One of those comments was genuinely useful feedback from Argentina about localization. That alone made it worth posting. But the post was gone from page 1 in what felt like minutes.
What's interesting is this isn't a weekend vibe-coded project - it involves actual physical production, printing, and shipping. But from the outside it probably looks like "another AI wrapper," which I think is the core problem: the flood of low-effort AI projects has made people reflexively skeptical of anything that mentions generation, even when there's real infrastructure behind it.
Hi from Germany! Thanks for the feedback.
You're right - the showcase sorting prioritizes "Real Books" (from customers who opted in to share theirs) over "Inspiration" (demo books we generated), which pushes the Spanish example further down. We'll fix the sorting so your language appears first regardless of type.
Good catch on the German book with the English idea - that's the customer's original input, which we didn't translate for the showcase. Will fix.
On "samples" vs "ideas": agreed it's confusing. We'll either make the distinction clearer or merge them into one gallery.
Thanks for the thoughtful feedback!
Posted by
storystarling on Sunday February 15, 2026 @06:48AM
from the show-hn dept.
Hi HN, I built StoryStarling. You describe a story idea and it generates a complete illustrated children's book, then prints and ships it.Every story, character, and illustration is generated from scratch, not a template with a name swapped in. Most books are about a new character your child can identify with: a raccoon starting kindergarten, a knight rescuing the Internet Connection, a...
yes, i had the same experience. As good as LLMs are now at coding - it seems they are still far away from being useful in vision dominated engineering tasks like CAD/design. I guess it is a training data problem. Maybe world models / artificial data can help here?
Nice work! We create personalized children's books - parents share their idea and photos, and AI brings their custom story to life with their child as the protagonist. We do hybrid fulfillment depending on the country. The PDF formatting challenges you mentioned are very real!
The story synopsis next to the preview gives you the full narrative arc before you commit. But fair point on wanting to see more.
You can edit or regenerate pages if something isn't working - it's iterative, not one-shot. Happy to help you try it out without payment - drop me an email.
Character consistency was the hardest problem, and honestly what took the longest to get right. We use reference images as style anchors, run multiple generation passes, and have an LLM "critic" that checks for visual inconsistencies and triggers regeneration when needed. It's not perfect but it's gotten to the point where parents are happy with the results.
On margins - tight but workable.
RTL = Right-to-Left languages - Arabic, Hebrew, Farsi, Urdu. The text rendering and page layout needs to flip for these, and it gets especially tricky with bilingual books where one language is RTL and the other is LTR.
What's the scenery? Happy to try it on our system if you want to share.
StoryStarling - Turn your story idea into a printed children's book
https://storystarling.com
Working on a platform where you describe a story concept and it becomes a real, illustrated picture book - professionally printed and shipped to your door.
The key difference from "personalized" book companies: this isn't template stories with a name swapped in. You bring an idea - maybe a book about a kid with a cochlear implant going to their first day of school, or a bilingual German-Turkish story about visiting grandma's village - and it generates a complete original narrative with consistent illustrations throughout.
You can upload reference photos so characters actually look like your child. Supports 30+ languages including bilingual editions on the same page.
Currently refining the showcase features and adding RTL language support.
Not AI. Anyway, the real issue is permissiveness vs strict parsing—real-world files are messy.
Raw CUDA works for the heavy lifting but I suspect it gets messy once you implement things like grammar constraints or beam search. You end up with complex state machines during inference and having standard library abstractions seems pretty important to keep that logic from becoming unmaintainable.
I suspect it is actually about maintaining permissiveness for malformed inputs rather than keeping security bugs. I ran into this building ingestion for a print-on-demand service where users upload technically broken PDFs that legacy viewers handle fine. If the new parser is stricter than the old one you end up rejecting files that used to work, which is a non-starter for the product.
The WASM constraints make sense given the resource limits, especially for mobile. If you are moving that compute server-side though I am curious about the unit economics. LaTeX pipelines are surprisingly heavy and I wonder how you manage the margins on that infrastructure at scale.
I thought the RTX 6000 Ada was 48GB? If you have 96GB available that implies a dual setup, so you must be relying on tensor parallelism to shard the model weights across the pair.
That sounds plausible. I've had to implement similar reputation-based limits on my own backend just to keep inference costs from exploding, so I sympathize with the fraud prevention angle. Masking that as a generic quota issue to push an upsell is pretty hostile though.
Fair point. I suspect it comes down to ghost reservations or stale caches. If a previous upload failed mid-flight but didn't roll back the quota reservation immediately, the backend thinks you're over the limit until a TTL expires. Or you delete something to free up space, but the decrement hasn't propagated to the replica checking your quota yet.
Do you have specific t/s numbers for those dense models? I'm curious just how severe the memory bandwidth bottleneck gets in practice.
I'm not sure I agree on the cost aspect though. For high-volume production workloads the API bills scale linearly and can get painful fast. If you can amortize the hardware over a year and keep the data local for privacy, the math often works out in favor of self-hosting.