Most image models were trained to read tags — woman, red dress, studio light, 8k, masterpiece. Krea 2 is different. Its text encoder is a Qwen3-VL language model, which means it reads your prompt the way a person reads a sentence. If you prompt Krea 2 like it’s an older tag-based model, you’re leaving most of its quality on the table.
Rule #1: write sentences, not tags
Krea 2 rewards natural prose. Instead of a comma salad, describe the scene as if you were briefing a photographer.
Weaker (tag style):
woman, denim jacket, city street, golden hour, bokeh, 50mm
Stronger (prose style):
A candid photo of a woman in a light denim jacket walking down a quiet city street at golden hour. Warm side lighting rakes across her face, the background is softly out of focus, shot on a 50mm lens at eye level.
The second prompt gives the LLM encoder a coherent situation to render — subject, wardrobe, environment, time of day, lighting direction, framing, and lens — and Krea 2 composes it far more faithfully.
What to include in a Krea 2 prompt
A reliable structure to describe, in order:
- Subject & action — who/what, and what they’re doing.
- Wardrobe / materials — specific fabrics, colours, textures.
- Pose & framing — full body, close-up, three-quarter, eye level, low angle.
- Lighting — direction and quality (soft window light, hard sun, warm practicals).
- Environment — where it is, and what’s in the background.
- Look & palette — the photographic style and dominant colours.
Concrete, physical detail works. If a specific look matters, spell it out in a full clause rather than hoping a single keyword carries it — the language encoder genuinely uses the extra description.
Rule #2: on Turbo, forget the negative prompt
Here’s the part that surprises people. Krea 2 Turbo runs at CFG 1, and at CFG 1 a classic negative prompt has essentially no effect — there’s no guidance gap for it to push against. Techniques that inject negatives at CFG 1 on other architectures (like NAG) don’t hook into Krea 2’s attention blocks, so they don’t help here either.
The practical takeaway: say everything in the positive prompt. Don’t rely on a negative like “no blur, no extra fingers.” Instead, positively describe the thing you want — “sharp focus throughout,” “hands resting naturally at her sides.” If you genuinely need strong negative guidance, that’s a reason to use the slower Raw variant at a higher CFG, where negatives regain some effect (at 2× the time and some risk of over-saturated colour).
Rule #3: lead with your trigger word (for LoRAs)
If you’re using a character or style LoRA, put its trigger word at the start of the prompt, then continue in prose. On AIFLUX this is handled for you — when you select a trained model, its trigger word is prepended automatically, so you just write the scene.
Try it live
The fastest way to feel the difference is to run the same idea as tags and as prose side by side. You can do that right now — open the Krea 2 generator on AIFLUX, no install required.
Frequently asked questions
Does Krea 2 use tags or natural language? Natural language. Its text encoder is an LLM (Qwen3-VL), so full sentences outperform tag lists.
Why doesn’t my negative prompt work in Krea 2? Krea 2 Turbo runs at CFG 1, where negative prompts have effectively no influence, and CFG-1 negative techniques like NAG don’t attach to this architecture. Describe what you want positively, or switch to Raw at a higher CFG.
How long should a Krea 2 prompt be? As long as it needs to describe the scene clearly. A few well-formed sentences usually beat both a single keyword and an overloaded wall of text.
Where do I put a LoRA trigger word? At the very start of the prompt. On AIFLUX it’s added automatically when you pick a model.
Put it into practice: Generate with Krea 2 on AIFLUX →