CookedThumb, and the bill that arrives with every generation
Building a product on a generative model changes the rules: every time a user hits the button, you pay. Here is how we dealt with that.
CookedThumb is a web app that turns screenshots and raw images into YouTube thumbnails and social assets, using a generative image model for composition, subject integration and visual styling. It generates multiple variants in parallel at 16:9, 9:16 and 1:1. It is built by MAXYMIZE.
Visit the CookedThumb website Generate thumbnails for YouTube, Shorts and social from your own images.
The problem it solves
The thumbnail decides whether a video gets opened. Anyone publishing regularly knows this, and also knows that producing a decent one takes either design skill or money, every time, for every piece of content.
Generic image generation tools do not cut it. They can produce an image, but they do not know that in a thumbnail the subject has to stay recognisable at tiny sizes, that text has to hold contrast, and that every platform wants different proportions.
With CookedThumb we started from the use case: upload a reference image for style and a subject, and get several variants already in the formats you need, editable by writing what you want changed.
The decisions that mattered
Applications built on generative models carry constraints traditional software does not have. These are the ones that shaped the product.
Pricing gets designed alongside the software
Every generation carries a cost paid to the model provider. In traditional software one more user costs almost nothing; here you pay per click. That shapes everything: whether a free tier can exist and how wide, where authentication becomes necessary, when credits start making sense. We made those calls alongside the technical ones, not afterwards.
Parallel generation with partial results
Variants all start together, because running them in sequence would make the wait unbearable. That forces you to handle things going wrong: if one generation fails or times out, the others still have to reach the user. Treating five independent requests as a single block would mean throwing away work already paid for.
Every aspect ratio is requested from the model
The shortcut would be deriving the vertical format by cropping the horizontal one. What you get is subjects cut in half and text running off frame. We ask the model for each format so it composes the scene for that space. It takes more generations and produces images people can actually use.
Prompts live in the repository, with their history
Output quality depends more on how the request is built than on which model you pick: structure, instruction order, handling of visual references. All of it lives in the codebase and is versioned. Editing a prompt changes how the product behaves for every user, so we treat it like any other release.
Files in object storage, references in the database
Generated images go into object storage while the database keeps references and metadata. The archive can grow without slowing down queries, and files are served directly instead of passing through application code.
What it is built with
A stack designed to handle slow, expensive operations without freezing the interface or producing surprises on the invoice.
Interface
- React 19
- TypeScript
- Tailwind CSS
- Vite
Artificial intelligence
- Generative image model
- Image-to-image editing
- Palette extraction
- Versioned prompts
Backend
- Serverless functions
- Serverless PostgreSQL
- Object storage
Accounts and payments
- Managed authentication
- Online payment gateway
- Signed tokens
- Credit system
What this means for your project
If you are considering a product that uses generative AI, these are worth knowing before you start.
- Products built on generative models, where per-operation cost is factored in alongside the pricing model instead of being discovered after launch.
- Slow, expensive operations handled without freezing the interface, with partial results, retries limited to what actually failed, and readable state while the user waits.
- Prompt engineering treated as part of the codebase: versioned, comparable, changeable without breaking what users have come to rely on.
- Integration between generation, file storage, accounts and payments. This is where most AI prototypes stop.
Frequently asked questions
What is CookedThumb?
CookedThumb is a web app that generates YouTube thumbnails and social assets from screenshots and raw images using a generative AI model. It produces several variants in parallel at 16:9, 9:16 and 1:1, with follow-up editing in plain language. It is built by MAXYMIZE.
Why can a generative AI app not be unlimited and free?
Because every generated image costs money paid to the model provider. In traditional software, serving one more user costs almost nothing; here the cost repeats on every operation. Credits, limits and plans come from that.
What does building something similar cost?
It depends on scope, but a product integrating generation, storage, accounts and payments sits in the range of a complete MVP. Our how we work page covers timelines and indicative price ranges. The heaviest part is not the AI itself, it is everything around it.
Can you add generative AI to our existing product?
Yes, and it is often the more sensible route: instead of building a new product you add a generative feature where it produces measurable value, with costs under control and a clear line between what the model generates and what stays deterministic.
Want to bring generative AI into your product?
Before choosing a model, the thing to work out is where AI gives you an advantage you can measure. That is where we start.