The Ultimate LoRA Training Guide for KREA2

Krea 2 is out, and it's already reshuffling how people think about custom training. A 12B-parameter diffusion transformer trained from scratch on real photography, no synthetic data in the mix, it's landed near the top of the independent image-model leaderboards and it ships as open weights. Naturally, the community wants LoRAs.

Here's the part that trips people up: Krea 2 isn't one model, it's two. There's Raw and there's Turbo, and if you don't understand the relationship between them, your first training run is going to go sideways.

1. Raw Trains, Turbo Runs

Krea 2 ships as a pair of checkpoints built to work together:

• Krea 2 Raw: the undistilled mid-training checkpoint. No CFG distillation, no aesthetic post-training applied yet. It's a wide-open, highly malleable latent space, which is exactly what you want when you're trying to push a model toward a new subject or style.
• Krea 2 Turbo: the distilled, post-trained sibling. Eight steps, no CFG, tuned for speed and out-of-the-box photographic quality.

The workflow that actually works: train your LoRA on Raw, then run it on Turbo. Turbo has already been aligned and distilled, which makes it stubborn during fine-tuning, post-trained models resist being pushed around. Raw hasn't been through that process, so your training has real room to move the model toward what you're teaching it. The adapter compatibility between the two is intentional; Krea built them as a matched pair specifically so this handoff works.

2. Captioning: Krea 2 Reads Sentences, Not Tag Soup

Krea 2's text understanding runs through a modern vision-language backbone (Qwen3-VL in most open training pipelines), and like the rest of that generation of models, it wants natural language, not a comma-separated pile of booru tags.

This means:

❌ woman, outdoor, red_dress, smiling, tree, daylight

✅ "A woman in a red dress smiling near a tree, photographed in natural daylight.”

Auto-captioning with a VLM gets you most of the way there quickly but don't stop there, though. Captions are what tell the trainer what to learn and what to ignore, if you don't call out a detail you don't want baked in, the model will happily absorb it as part of the concept.

Auto-captioners are notoriously verbose. They love to state the obvious, and this is where you need to step in and edit.

The 30-50 Word Sweet Spot

When editing your generated text files, ruthlessly cut them down. Your target length is 30 to 50 words. Absolutely do not exceed 70 words.

When a caption gets too long, it becomes "word salad." The text encoder's attention mechanism gets diluted across too many tokens, making it harder for the model to isolate the specific concept you are trying to teach it.

What to Cut from Auto-Captions

To keep your captions lean and focused, delete the following:

• Meta-descriptions: Auto-captioners love to start with "The overall image shows..." or "This is a photograph of..." Cross these out immediately. Start directly with the subject.
• Redundant background details: If the background isn't part of the concept you are training, don't over-describe it. "Outdoors in a city" is fine; you don't need to describe every window on the buildings.
• Lighting descriptions (Usually): Unless you are specifically building a lighting LoRA (e.g., a "Cinematic Neon Noir" style), remove phrases like "soft lighting" or "harsh shadows." If you caption the lighting, the model associates that lighting only with your prompt, rather than learning it as a flexible part of the subject. Let the base model handle the lighting during generation.

3. The Truth About Trigger Words

In the SD 1.5 and SDXL eras, trainers loved using rare, gibberish tokens as triggers (e.g., th1sch9r or zqx_style). The logic was sound: use a blank token that has no prior meaning so the model doesn't get confused.

Do not do this with KREA2. In KREA2 training Random strings are either ignored entirely or learned incredibly poorly, resulting in a trigger words showing as text or watermark.

Rule 1: Use Natural, Readable Words

If you need a trigger word, combine natural, aesthetically pleasing words. Underscores and alphanumeric mashups are out.

• Instead of: lum1n9
• Use: lumina portrait
• Instead of: zyra_style
• Use: zyra outfit

Rule 2: Make the Trigger Part of the Sentence

Do not just slap your trigger word at the beginning of the text file followed by a comma. It must be grammatically integrated into the caption.

Examples:

Rule 3: Triggers Aren't Always Required

Look at your caption dataset. If certain terms or phrases (like "vintage anime" or "oil painting") naturally repeat in 70% or more of your captions, the model will already learn to associate your style with those existing words. For Style LoRAs, you almost never need custom triggers.

Rule 4: Absolute Consistency

If you do choose to use a trigger word, it must appear in the every single image caption otherwise the model will struggle to bind the concept to the word.

4. Understanding Exposure

Before we get into the exact settings for different types of LoRAs, you need to understand the most important metric in LoRA training: Exposures per image.

Forget about "Total Steps" for a moment. Total steps are just a byproduct of your dataset size. What really matters is how many times the model looks at a single image during the entire training run.

Here is the formula: Images × Repeats × Epochs = Total Exposures (and Total Steps)

If you have 100 images, set to 4 repeats, running for 10 epochs:

• 100 × 4 × 10 = 4000 Total Steps
• In this scenario, the model looks at each individual image 40 times. (4 repeats × 10 epochs = 40 Exposures per image).

The Golden Rule of Exposure: If you want a LoRA to train faster (in fewer total steps), do not decrease the exposure per image. Instead, lower the amount of images in your dataset. A 30-image dataset with 40 exposures will learn a face perfectly in 1,200 steps. A 100-image dataset with only 12 exposures will fail, even if it runs for the exact same 1,200 total steps.

5. Dataset Blueprint by LoRA Type

A. The Style LoRA Blueprint

A Style LoRA alters the artistic medium, rendering technique, or overall aesthetic (e.g., 90s dark fantasy, watercolor, claymation).

The goal here is to remove the model's dependency on specific objects and force it to learn the technique itself. To do this, your dataset must be highly diverse in subject matter (people, landscapes, animals, cars) but strictly unified in visual style.

• Images: 50 - 100 (High diversity of subjects required).
• Target Exposure per Image: 35 - 40
• Epochs: 8 - 10
• Total Steps: 2700 - 3000 (Upto 4000 if your if your dataset hits 100 images)

Pro-Tip for Styles: Do not use a custom trigger word. Instead, use a descriptive phrase that already exists in the model's vocabulary, like "A vintage comic book illustration of..." Put that exact phrase at the start of every caption.

B. The Concept & Clothing LoRA Blueprint

A Concept LoRA focuses on a specific item, outfit, vehicle, or repeating visual theme (e.g., a specific set of futuristic armor, or a distinct type of architecture).

The secret to a perfect concept LoRA is Isolation through Variation. If you are training a specific leather jacket, your dataset needs to feature that exact jacket worn by men, women, different ethnicities, in different locations, and at different times of day. If 80% of your dataset features a blonde woman wearing the jacket, the model will start turning everyone into a blonde woman the moment you prompt for the jacket. Vary everything except the concept.

• Images: 50 - 100
• Target Exposure per Image: 30 - 40
• Epochs: 8 - 10
• Total Steps: 2000 - 2500

Pro-Tip for Concepts: Use a natural trigger phrase integrated into the sentence. Example: "A person wearing a crimson vanguard jacket standing in an alleyway."

C. The Person & Character LoRA Blueprint

Training a specific human face or fictional character requires the smallest dataset, but the highest quality curation.

You need absolute consistency in the face, but variation in everything else. Provide a mix of extreme close-ups, medium portraits, and full-body shots. If you only provide portraits, the LoRA will completely break when a user tries to generate a wide-angle action shot. Vary the clothing, angles, and lighting, but avoid extreme distortions or heavy filters on the face.

• Images: 20 - 30
• Target Exposure per Image: 25 - 35
• Epochs: 8 - 10
• Total Steps: 1600 - 2000

Pro-Tip for Characters: In your captions, do not describe the permanent features of the character (e.g., don't describe their eye color or specific nose shape). Only describe the variables in the image (their pose, their clothing, the background). This forces the model to bind the permanent facial features directly to your trigger word.

To Summarize

KREA2 heavily rewards precision over sheer volume. It does not need thousands of images or endless training steps; it needs clean, grammatical language and thoughtfully structured datasets.By leaving booru tags behind, embracing conversational captions, and carefully managing your exposure rates, you will unlock the full potential of this incredible model.