The setup that breaks the habit
Most of us learned to split prompts into two boxes: positive for what you want, negative for what you don't. That habit works fine on models sampled with classifier-free guidance at normal strength. It falls apart on Krea 2.
Krea 2's turbo-style sampling runs at CFG 1.0. At that setting, the math that makes a negative prompt matter is effectively turned off. CFG is the knob that weighs the positive prediction against the negative one. At 1.0, the model is just following the positive prompt on its own. Nothing you type in the negative field changes the output.
This isn't a bug in your workflow. It's how the sampler is meant to run. Krea 2 is built for fast, low-step generation, and CFG 1 is part of what makes that speed possible. The tradeoff is that an entire category of control — negative prompting — stops working.
Why this bites people specifically on logos and text
The most common place this shows up: someone adds "no logos, no text on clothing, no emblems" to their negative prompt, generates a batch, and still gets shirts with random brand marks or garbled text patches. They assume the prompt is being ignored or the model is stubborn. It's neither. The negative prompt was never read in the first place.
This matters more for character work than for one-off art pieces. If you're generating a character across many outfits, scenes, or batches, stray logos and text are the kind of detail that makes images unusable for anything beyond a personal test. You don't want to catch it on inspection after a 50-image run.
Where the exclusion actually has to live
If CFG 1 kills the negative prompt, the only place left to steer the model away from something is the positive prompt itself. You state the constraint as part of what you want, not what you don't want.
Instead of relying on a negative field, something like a short clause stating the clothing should be plain, without logos, emblems, or visible text, works because the model treats it as a positive instruction to follow, not a negative one to subtract. It's a small phrasing shift, but it's the difference between a clause that does nothing and one that actually constrains the output.
A few other categories where this trick helps:
- Branding and text on packaging or signage in the background
- Visible watermarks or artifact-style text the model sometimes hallucinates
- Specific objects you want absent from a scene (describe the scene as not having them, in positive terms)
The pattern is the same each time: say what should be true of the image, not what shouldn't appear in it.
Why we bake this into the templates
We got tired of re-explaining this every time someone asked why their negative prompt wasn't doing anything on Krea 2. So we added a short no-logos clause to every Gmanski Krea 2 template by default. It's a few words, sitting quietly in the positive prompt, doing the job the negative field can't.
If you're building your own prompts from scratch rather than starting from a template, this is the one habit worth carrying over: whenever you catch yourself reaching for the negative field on a Krea 2 run, stop and ask whether that exclusion can be rephrased as a positive instruction instead. Almost always, it can.
This doesn't just apply to clothing. Any workflow doing batch character work — swapping outfits, scenes, or poses across a large set — benefits from the same fix, since the error compounds across every image in the batch rather than showing up once.
Checking your own setup
If you're not sure whether this applies to a given workflow, look at the CFG value on the sampler node. If it's set to 1.0, assume the negative prompt box is inert and move anything important into the positive prompt. If CFG is higher, the negative prompt is doing real work and you can keep using it as normal.
This is also a good moment to check any saved presets or node groups you've been reusing from older, higher-CFG workflows. Copying a negative prompt forward into a Krea 2 setup without noticing the CFG change is an easy way to end up with a quiet, invisible mistake baked into every batch.
Takeaway
On Krea 2's CFG 1.0 sampling, the negative prompt field does nothing — move exclusions like no-logos or no-text clauses into the positive prompt, where the model can actually act on them. The Krea 2 workflow ships with this clause already in place, and the bulk character swap workflow carries it through batch runs so you're not fixing it image by image.
