News

Doing everything was the problem

CogniWerk·7 Aug 2026

The most useful sentence anyone said about our product this year came from someone who actually used it. Not from a demo, not from a first call, but after weeks of daily work. Paraphrased: the platform allows everything, and that is exactly what makes it hard to manage when you are looking for one specific solution.

We chewed on that one for a while, because “allows everything” was what we had built, deliberately. 20 leading image models in one place. Worklets for recurring production tasks. Video, LoRAs, batch processing, granular control over every parameter. A toolbox holding everything a marketing team running campaigns could need.

Except nobody arrives wanting a toolbox. People arrive with a job. The product shot needs a different background. The bottle in the image should be milky rather than transparent. One motif needs six variants that look like they came from the same shoot. Turn up with a job like that and find your first decision is which of twenty models fits it, and you have not started the job yet, you have added one.

For a long time we filed that under learning curve. Every piece of software has one. But a learning curve assumes somebody sets foot on it, and the more feedback we gathered, the clearer it became that the effort was sitting in the wrong place. Not in the work, but in front of it.

What changed

So we rebuilt the platform around what people are actually trying to do, rather than around everything the system is technically capable of.

Use cases instead of model selection. CogniHub is now split into clear areas: recent, base models, use cases, video models, LoRAs. If you came to produce something, you go to use cases, and what you find there are not model names but jobs: replace clothing, modify a product, generate consistent shots, relight a product, place a product in a scene. Each card says in one line what it does.

The CogniHub panel with tabs for recent, base models, use cases, video models and LoRAs, showing five use-case cards with their credit cost and input type. — Use cases in CogniHub: jobs rather than model names, each card explained in a single line.

Costs stated up front. Every card shows how many credits a run costs, and whether the tool works from text or from your own images. No surprise after the click, no mental arithmetic.

Five base model cards showing credit costs from 5 to 10 credits and labels for text or image input. — The same on base models: credits per run and input type stated on the card, before anything is clicked.

A guided path instead of an empty form. Use cases now run in steps. Three inputs, a progress bar, one field at a time, with back and skip where skipping makes sense. At the end there is a summary of image, change and subject, reviewable before any credits are spent. And once you know the flow, one checkbox turns it off for that tool for good.

Guided step with progress indicator, an upload field for the product image and a next button. — Step one of three: one required field, a progress bar, and one line explaining what is needed here.

Review summary listing product image, modification and subject, with a checkbox to skip the walkthrough next time. — Before generating: everything shown back to you, plus the checkbox that skips the flow from here on.

Help exactly where it is needed. The modification field does not say “describe your change.” It says that “make the bottle matte black” produces better results than “make it nicer.” The subject field explains what it is actually for: name the product so it stays anchored while the change is applied, and leave it blank if you would rather have the main product found for you. That is the kind of knowledge usually delivered in an onboarding call and forgotten three days later. Now it sits where the typing happens.

Modification input field with a hint that a concrete instruction works better than a vague one. — The guidance sits in the field itself, not in documentation nobody reads.

Optional subject field explaining that naming the product anchors it, and that it can be left blank. — Second example: the subject field explains what it does and what happens if you leave it empty.

What stayed the same

The depth has not gone anywhere, it simply is not the front door any more. Anyone who wants full control over models, parameters and seeds finds it unchanged in the same place. Base models remain directly accessible, as do batch runs and variant runs. And the guided flow is an offer rather than an obligation: one checkbox and it is out of the way for that tool permanently. The difference is that nobody has to walk through the toolbox to reach the workbench.

And the ground underneath is unchanged: EU-hosted, GDPR-compliant, aligned with the EU AI Act. Nothing moves here, because nothing here needed improving.

Two cleaning product bottles, the spray bottle now orange and the concentrate bottle milky grey, both labels still legible. — Two changes in a single run, with labels and small print preserved intact.

Why we are writing this so plainly

We could have written this as a feature announcement without conceding there was a real problem before. We think that is the worse version. The feedback that helped us most this year was the uncomfortable kind: too raw, not intuitive enough, the results do not yet match our requirements. You only get sentences like that from people who believe they will land somewhere.

So, directly: if you tested CogniWerk and quietly put it down, we would like to know why. Twenty minutes, no pitch, no demo. Criticism is more useful to us than agreement, and this update is the best evidence of that we can offer.

And if you would rather see for yourself, the use cases are live. Bring a brief that is already sitting on your desk.