What GPT-Image-1 Changed, and What Kitana Uses Instead

Model names rarely deserve an article. This one is still worth understanding, not because Kitana runs it (it does not, and the second half of this piece says what does) but because it marked a shift in what image models are optimised for, and that shift explains why the apps built on them changed shape.

The first half describes the picture as of early 2026. Model releases move fast enough that specifics date quickly; the shift is the durable part.

What it was, and what it changed

GPT-Image-1 is OpenAI's image generation and editing model, made available through its API in 2025. Two characteristics stood out, and both are about precision rather than beauty.

Instruction following. It mapped the words in a prompt onto the output more reliably than the generation before it. Specific relational instructions, put this here, change only that, had a better chance of producing the specific change rather than a generally similar picture.

Text rendering. Markedly better than what came before, which had been uniformly poor since 2022.

Neither is "makes prettier images". Both are about doing what you said, and within a year every competing model was being judged on the same two things.

Why that shift matters

Early image models were optimised, implicitly, for striking output. You described a vibe, you got something impressive, and you had limited ability to steer it. Prompting culture reflected that: long strings of quality adjectives, style names and modifiers, because that was the available lever.

A model that follows instructions inverts the incentive. Now the useful prompt is the precise one, and the adjective pile is actively counterproductive; it consumes attention that a describable detail could use.

This is why prompting advice from 2023 often makes results worse on current models. The advice was correct for the models it was written for.

What it changed for apps

The practical consequence, and the reason this is visible in consumer products rather than only in research.

Editing became as reliable as generating. A model that follows instructions can be asked to change one thing in an existing image and leave the rest, which is what most people actually want from a photo app. Generating a new image from nothing is a smaller market than improving a photo you already have.

Prompts could be narrowed on your behalf. Well-built apps translate a button press into a precise instruction rather than passing your words through. That works far better when the model reliably obeys precise instructions, and it is a large part of why two apps on the same model produce different results.

Text stopped being an automatic disqualifier for a category of use. Still unreliable for sentences, usable for short words.

What Kitana actually runs

Honesty about this is the point of a trust article, so here is the current picture.

Kitana used GPT-Image-1 as one of two providers for part of 2026, and removed it in September. Every image and every clip is now produced through IVCreator, a provider that runs a set of third-party models and lets the app pick one per tool: one model handles text-to-image, others handle photo edits and the photo-to-video clip. The Terms of Use and the Privacy Policy name IVCreator as the provider, and there is no OpenAI call anywhere in the app or the server.

Two things drove the change. On the tasks Kitana actually does, which are mostly edits of a photo you upload rather than images from nothing, the results were no longer better than the alternative. And running two providers meant two credentials, two failure modes and two bills for one feature. One provider, held only on the server, is simpler to secure and simpler to reason about; the reasons that matters are in how AI photo apps protect your data.

What did not change is the lesson. The shift GPT-Image-1 marked, toward models that obey a precise instruction, is now the baseline every provider competes on, and the models behind Kitana are judged by the same standard: does "keep the face, change the background" do exactly that.

How much should the model name affect your choice?

Less than the marketing implies.

Most consumer photo apps call one of a small number of model providers. Raw capability is therefore more similar across the market than it appears. What genuinely differs is everything around the model: instruction framing, image pre-processing, failure handling, data practices, the free tier. None of that is visible from a model name, and an app that leads with one is telling you about its supplier, not about its results.

The test that answers the real question takes ten minutes and uses your own photo: how to choose the right AI photo app.

What did not change

Worth stating, because every release brings claims that the limits are gone.

• Hands improved and remain the most common tell.
• Plausible is still not accurate. The model produces something consistent with its training, never something verified.
• Counting and spatial relations are followed more closely and still loosely.
• Nothing is retrieved. There is no image database; everything is generated.

These follow from how diffusion models work rather than from any particular model's quality, which is why successive generations move them without removing them. The mechanism is in what is text-to-image AI, and the arc these releases sit within is in the history of AI image generation.

The practical takeaway

If your prompts read like a list of quality adjectives, rewrite them as descriptions of what should be visible. That single change does more for your results on any current model, whichever provider is behind the button, than switching apps.

Try it: write one descriptive prompt in the browser text-to-image tool, then test the "keep the face, change the background" instruction on your own photo with background replacement.