TaleByYou

Started with Stable Diffusion and Midjourney. The character looked completely different on every page - different hair, different face, sometimes different race. Parents would never buy books like that.

Spent weeks testing different approaches. Finally found Leonardo AI's Character Reference feature (ControlNet preprocessor 133). You upload one photo and it uses that as reference for every single generation. The AI actually maintains the same facial features, hair style, everything.

Now the character looks consistent across all 20+ pages. Parents send me photos of their kids recognizing themselves in the books. Way less regenerations needed.

Had a mandatory registration screen. 60% of people bounced immediately. Makes sense - who wants to create an account before knowing if the product is even good?

Built a guest session system with encrypted cookies. You can create up to 2 books without any account. Session lasts 30 days. If you order, you get an option to register and claim your books. Prevents abuse but lets people try it first.

Conversion went up 35%. People make their first book, see that it actually works, then buy. Most register after they see the final product.

Lulu rejected 40% of my PDFs. They need exact dimensions with bleed margins, spine width that changes based on page count, safety zones, specific color profiles. The error messages were useless - just 'PDF validation failed.'

Used their API to calculate cover dimensions dynamically. Built a PDF generator with PDFKit where I control every pixel. Added my own validation before sending to Lulu. Tested with books of different lengths until I got it right.

99.8% success rate now. Only fails if someone uploads a corrupted image. Saves me hours of back-and-forth with customers about failed print jobs.

The total price depends on: printing cost (changes with page count), shipping (5 options, varies by country), AI generation costs, and my profit margin. All of this needs to calculate in real-time before someone checks out.

Chained API calls in a service layer: get printing cost from Lulu, get shipping for their location, sum up AI costs from my database, apply markup, return breakdown. Cache the result so I'm not hammering Lulu's API for every price check.

Instant pricing. Customers see the breakdown and trust it. I can adjust margins without updating prices manually.

Each image generation takes 20-60 seconds. Started with a loading spinner that blocked the whole page. People would close the tab, lose their progress, and email me confused.

Switched to webhooks. Start the generation, let people close the page. Leonardo pings me when it's done, I send them an email with a link back. If they stay on the page, they see progress. Either way, nothing gets lost.

Way better UX. People start the generation and go do something else. Come back when they get the email. No more lost work.

Storing PDFs permanently for every order. 100 orders = 200 files at 5MB each = 1GB monthly. Costs kept climbing and most people download once then never again.

Generate PDFs on-demand when someone orders. Upload to S3 with 7-day expiration. They get a download link that works for a week. If they lose it, they can request a new one (charge a small fee to prevent abuse).

Cut storage costs by 90%. People download within a day anyway. Less than 2% ask for regeneration.

Users pick from 3 AI variations per page, across 20+ pages. My first version: one page at a time. Took forever. Thumbnails were too small to actually see details.

Built a grid showing all pages. Click for full-size lightbox. Arrow keys to navigate. Checkmark shows what's selected. Added bulk actions to auto-select for all pages. Keyboard shortcuts for people who want to go fast.

Selection time dropped from 10 minutes to 2. People actually enjoy this part now.