Every image version is now on the page
A friend of ours laughed at our first fish soup photo. Fairly: the machine had served up a bowl of raw fish and announced, with total confidence, that this was halászlé. It wasn't a plated dish so much as the opening scene of a horror film.
We could have swapped in a prettier picture and pretended it never happened. Instead the feedback became the next feature.
What changed
Under every recipe you'll now find the full gallery — not just the best photo, but every version from the first attempt to the current one. And each one says exactly what produced it:
- The image model
- Which model drew it, at what precision, and how large its weights are in gigabytes. That last number isn't decoration — it's what a step up in quality actually costs.
- The text encoder
- This is the model that reads the description, and it decides what ends up in the picture — not merely how pretty it looks. The raw fish was decided here, not at the drawing stage.
- The prompt and settings
- The full instruction, the seed and the rest of the settings open in one tap. We're not hiding them: if you want to see what made the fourth image better than the second, it's right there.
- The licence
- Every image shows the licence it was produced under, and we only offer a download link where that model's licence genuinely allows it. We don't want anyone put in breach by a file they took from us.
The images are shown in 4:3, because that is the shape the model works in. Cropping them to 16:9 would throw away a quarter of the frame — exactly the part we're asking you to compare.
The bad picture is part of the story
We didn't delete the first version. It's still there, first in the row, exactly as it was. On a reference project about AI, showing the progress is more honest than showing only the result — and frankly the raw-fish photo tells you more about where these models actually are today than any polished marketing shot would.
From feedback to live: hours
Here's the part worth noticing. From that laugh to this feature running in production — database migration, admin interface, public gallery and licence handling included — took hours, not weeks. Not because we rushed it, but because on the ATAILA platform the whole path — code, build, test, release, operations — runs in one place, on our own hardware. An idea can go live the same day it lands.
That is what OpenChef is really demonstrating. Not the recipes — how quickly a piece of feedback becomes a working product.
Any questions?
We're happy to answer anything — what each model can and can't do, why we picked these ones, how much hardware they need, or what all of this would look like on your own project. Reach us through the contact page. And if one of the pictures still looks ridiculous, tell us that too — so far that's exactly what has moved this project forward.
Interested in something like this?
We can build it for you. One workflow, turned into a private, AI-powered app that runs on European hardware you control — your AI, your data, your hardware. OpenChef is what that looks like when the subject is recipes; yours would look like your business.
You imagine it. We make it real.