Train a Style Control With Six Sentences and Zero Images
A published recipe makes a reusable style dial out of prompt pairs. No dataset, no scraping, no rights question. Here's the method, the cost, and the four places it falls apart.
Every custom style tool you have ever trained started the same way: you went and got pictures.
Twenty of them, or two hundred, or two thousand. You cropped them, captioned them, and at some point you had a quiet conversation with yourself about where they came from and whether you were allowed to use them that way. That gathering step is the real cost of custom style work. The training run is an afternoon. The dataset is the part that takes weeks and carries the risk.
A recipe published this week skips it entirely. Not "uses fewer images." Uses none.
The releases are called particle sliders, posted early Wednesday, September 23 by ntc-ai for two different image models, with the code and reproduction notes public. The output is a small add-on file you load like any other, and it behaves like a dial rather than a preset: turn it to zero and it does nothing, turn it to one and a plain blue service robot grows angular shoulder armour and weathered panels in the same yard, from the same prompt, on the same seed.
What makes it worth your afternoon is not the effect. It's how the style got defined.
The style is the difference between two sentences
Here is the whole idea.
You write six pairs of prompts. Each pair says the same thing twice, once plain and once in the style you want. Plain: "a warrior in steel armour, standing guard." Styled: the same warrior, but described as the imposing end-of-game version, with the heavy silhouette and the oversized weapon.
The model renders both sides of each pair itself. It already knows how to make both. Then the training run watches the difference between them, and teaches a small add-on file to push the plain version toward the styled one without being told to in words. When it's done, the add-on carries the difference. Apply it to a prompt that says nothing about armour and you get the armour anyway.
That's the trick. The style is not extracted from a folder of pictures. It's extracted from the gap between two descriptions the model can already follow. You are not teaching the model anything new. You are isolating a direction it already knows how to move in, and putting a handle on it.
Which is why there's no dataset. There is nothing to collect, nothing to caption, nothing to clear, and no question about whose work is inside the result. For anyone who has stalled a project over exactly that question, this is the part that matters.
The bleed problem, and the fix
The obvious failure of this approach is that the effect spreads. You train a dial that makes warriors imposing and suddenly every bowl of fruit in every render looks imposing too.
The recipe handles this with something small and clever. Every fifth training update, instead of a warrior pair, it uses a plain unrelated prompt: a lake, a bicycle, a cat. On those updates the job is the opposite of learning. The job is staying identical. The add-on gets pulled toward the style four times, then held still once, over and over.
The authors publish a fruit test in every release for exactly this reason. Same bowl of fruit, dial off and dial on, so you can see how much bleed survived. On the Krea 2 release they say plainly that the fruit shows "some rendering-style drift." That's the honest version of a number nobody else publishes.
What it costs
The headline is two minutes, and the authors are careful about what that means, so I will be too.
On the smaller model, 1,600 training updates took 134.4 seconds on one rented RTX A6000, including saving the file and rendering a first preview, with a peak memory use of 1,541 MiB. That is a small card's worth of memory and about the length of a coffee run. The timing evidence is published as a JSON file you can go read.
But that's the fast run, not the finished one. The same release also did a long run that settled at step 28,000 and took about 47.5 minutes including checks and sample renders. The authors state outright that the two-minute figure describes a separate short run and is not a convergence claim, and they ship both versions so you can compare.
The file you end up with on the small model is 6.8 MB, with a compressed half-size version at 3.4 MB. On Krea 2 the files are larger, about 77 MB with a 38 MB compressed version, because it is a much bigger model.
For the Krea 2 release the published recipe is 400 updates at 512 pixels with the six caption pairs and the every-fifth-update preservation. That is a small enough job that the barrier stops being money and starts being whether you can write six good pairs of sentences.
Put this into practice
Start by running someone else's dial before you train your own. It takes five minutes and tells you whether the effect is what you think it is.
If you work in ComfyUI, go to the Krea 2 sliders and take the ComfyUI export of either "Final Boss" or "Eldritch." Load it with the stock Load LoRA node. Set model strength to 1 and text strength to 0. Start from the Krea 2 Turbo template at its usual eight-step settings. No custom node is needed. Render the same prompt with strength 0 and strength 1 and put them side by side.
If you want to train one, clone the repository, install the requirements, and run the training script in its dry-run mode first to confirm the setup works before you rent anything.
Then the actual work, which is the writing. Six pairs. My advice on making them good:
Change one axis, not three. If your styled prompt is heavier, darker and more ornate than the plain one, the dial learns a blur of all three and you can't separate them later. Pick the single quality you want on a handle.
Keep the subject fixed inside each pair. The pair should differ only in the style words. If the plain one says "a knight" and the styled one says "a demon lord," the dial learns to change the subject, not the treatment.
Spread the six across different subject types. A person, an object, a vehicle, a creature, an interior, a wide exterior. Six variations on knights gives you a knight dial.
Write the plain side genuinely plain. The gap is the lesson. A plain prompt that already leans stylish gives the dial a small gap to learn from and a weak effect at the end.
Keep your own fruit test. One unrelated prompt you render at both strengths every time. Bleed is the failure mode and you want to see it early.
Where it breaks
This is the part the enthusiasm usually skips, and the authors do not skip it, which is why I trust the rest.
It can move things it shouldn't. The release says the original dial "can shift unrelated objects," and shows it. This is not a solved problem, it's a managed one. The preservation step reduces bleed and does not remove it.
The compressed version is weaker. Both releases ship a smaller half-size file alongside the full one, and both say the small one can alter details or weaken the effect. The authors publish approximation error tables and then say, correctly, that those numbers measure how close the copy is to the original, not whether either looks good. Compare the images. Don't compare the tables.
The speed number is one run. One card, one model, one measurement. It is not a benchmark and the authors say so on the front page. Your own run on a different model at a different size will not take 134 seconds, and the Krea 2 recipe is a different shape entirely.
The small model is small. Supra makes 256-pixel images. Faces, fine detail and lettering are poor, and the release notes that unwanted text can show up in cave images. It is a fine place to learn the recipe and a bad place to make deliverables. Krea 2 is the one to use for real work, and the Krea files carry the Krea 2 Community License, not an open one, so read the terms before selling anything made with them. The Supra pair is Apache 2.0. Two releases from the same author on the same day, two different licences, because the licence follows the base model rather than the add-on.
ComfyUI support is uneven between the two. The Krea 2 release ships a ComfyUI export and a setup guide. The Supra one says its ComfyUI compatibility has not been validated. Check which you're downloading.
The part worth thinking about
I keep coming back to what this does to the argument about training data.
Not solves. Does to. A dial trained this way still sits on top of a model trained on somebody's pictures, and nothing here changes that. What it changes is the layer you work at. When the custom style is defined as the distance between two sentences, the thing you made and the thing you own is the writing. The pairs are yours. The direction is yours. Six sentences is a small enough artifact to keep in a text file next to the project and hand to whoever picks it up after you.
Whether that holds up on bigger models with more specific styles is genuinely unknown, and the releases so far cover two models and a handful of effects. But the cost of finding out is now six sentences and a rented hour.
Write the six pairs before you rent anything. If you can't make the plain and styled versions feel clearly different in words, the dial won't be able to either.
Medium metadata
Title: Train a Style Control With Six Sentences and Zero Images Subtitle: A published recipe makes a reusable style dial out of prompt pairs. No dataset, no scraping, no rights question. Here's the method, the cost, and the four places it falls apart. Tags: Generative AI, AI Art, Creative Tools, ComfyUI, Machine Learning Est. read: 8 minutes