Krea 2 Complete Guide: Styles, References, Identity Editing

Krea 2 works best as a layered visual-control system, not a single prompt-to-image tool. Prompts control content, S-Refs control visual style, Moodboards shape the broader aesthetic, and identity workflows help preserve the subject while other elements change.

The challenge is choosing the right control for the right task. S-Refs cannot reliably lock a face, Moodboards are not identity tools, and combining style, pose, clothing, and character consistency in one prompt often causes drift. A stronger AI design workflow separates exploration, art direction, identity control, and staged editing.

For teams scaling this kind of multi-model creative workflow, Virse combines AI creation with a canvas-based design workspace. Paid plans include unlimited use of 40+ models, including Nano Banana 2 and GPT Image 2, with unlimited seats. New users also receive free credits, enough for up to 10 Nano Banana 2 images or one Seedance 2.0 video.

How Does Krea 2 Separate Styles, References, and Identity Control?

The most useful way to understand Krea 2 is to stop treating every image input as the same kind of reference. Krea itself frames image generation around two questions: what should appear, and how should it look? Its reference and Moodboard systems extend that separation into a more practical art-direction workflow.

From a product design perspective, the key distinction is simple: S-Ref answers “How should this look?” Identity control answers “Who or what must remain recognizable?” Confusing those two problems is one of the fastest ways to create inconsistent results.

How Should You Prompt Krea 2 for Better Style Control?

Start Short When the Creative Direction Is Still Open

Krea 2 is designed for visual exploration. A short prompt leaves room for the model to propose composition, color, camera treatment, and aesthetic direction before you lock the concept down.

A practical workflow is:

1. Define the subject and action.
2. Generate several visual directions.
3. Select the strongest direction.
4. Add camera, lighting, composition, material, and texture.
5. Add references once the aesthetic is easier to show than describe.

This reconciles two seemingly different prompting strategies: start vague when exploring, then become specific when producing. A structured AI design prompt workflow can help make that transition more deliberate.

Describe Visible Design Decisions, Not Generic Quality

Once the direction is clear, replace vague quality language with things a designer could actually see: low-angle framing, shallow depth of field, matte plastic, polished chrome, hard side light, film grain, motion blur, rough paper texture, or low dynamic range.

The goal is not a longer prompt. The goal is to describe only the visual decisions that still need explicit control.

How Do Krea 2 S-Refs and Style Strength Work?

What Krea 2 S-Ref Actually Transfers

A Krea 2 S-Ref transfers visual characteristics such as color, texture, composition cues, and the painted or photographic quality of an image to a different subject. It is therefore a style-control mechanism rather than an identity-locking system.

This is useful in campaign design. A team can preserve the same metallic treatment, lighting language, grain, or illustration character across eyewear, packaging, products, and editorial imagery without rewriting the entire aesthetic in every prompt.

Krea 2 S-Ref Strength: 20% vs 50% vs 80%

Krea's own examples make the trade-off unusually clear.

At about 20%, the reference leaves hints of palette or painted character while the model remains dominant. At 50%, Krea describes the result as a balanced blend. At 80%, the reference can take over palette, brushwork, and composition strongly enough to affect the subject itself.

That last behavior is style leakage. In production, I would start near 50%, compare several outputs, and increase strength only when style adherence is clearly too weak.

How Do You Mix Multiple Krea 2 S-Refs and Moodboards?

Use Multiple S-Refs as Separate Art-Direction Layers

Krea's standard Style Reference workflow supports up to four S-Refs, each with its own strength control.

Krea demonstrates combinations such as 70% + 50%, 75% + 57%, and 80% + 60% + 44%. The practical lesson is not to copy those numbers, but to assign each reference a visual responsibility.

For example:

• Reference A controls palette and material.
• Reference B controls lighting or photographic treatment.
• Reference C controls texture or illustration language.

Do not maximize every S-Ref at once. When several references compete strongly, it becomes harder to diagnose which image caused an unwanted material, composition, or content shift.

Krea 2 Moodboard vs S-Ref

A Moodboard is not simply a larger stack of S-Refs.

Krea says Moodboards can analyze larger image sets using style transfer plus custom LLM and clustering methods. The system can reason over concepts, recurring characters, expressions, compositions, atmosphere, and overall mood, then produce a Taste Profile, Keywords, and Avoids.

For a specific campaign treatment, use S-Refs. For a longer-running brand world, Moodboards are the stronger system-level tool.

How Do Krea 2 Generative Sliders Improve Creative Control?

Generative Sliders let designers change visual behavior while keeping the prompt stable. Krea currently exposes Intensity, Complexity, Movement, and Creativity.

Intensity moves toward stronger or quieter stylization. Complexity controls visual density. Movement affects camera and pose energy. Creativity determines how much Krea expands what you wrote.

For Creativity, Raw means no prompt expansion, Low stays closer to the prompt, Medium provides a balanced interpretation, and High gives Krea more freedom over style, mood, and aesthetics.

This Creativity setting called Raw should not be confused with Krea 2 Raw, the separate undistilled model checkpoint discussed below.

Krea 2 Medium vs Large vs Raw vs Turbo: Which Should You Use?

Krea 2 Medium vs Large

Krea positions Medium as the faster, cheaper, more heavily post-trained default with stable results across a broad range of tasks. Large is more than twice its size, has softer post-training, and produces a rawer, more textured, flexible character with particular strength in photorealism and experimental looks.

That means “Medium for art, Large for photography” is too simplistic. Medium is a strong default; Large makes more sense when you specifically value texture, photorealism, motion blur, grain, or wider expressive range.

Current Krea API pricing lists Medium at $0.030 per text-to-image generation and $0.035 with S-Refs, compared with $0.060 and $0.065 for Large.

Krea 2 Raw vs Turbo

Krea 2 Raw is a different model family context. Krea describes Raw as the undistilled base checkpoint for fine-tuning and LoRA training, while Turbo is an 8-step distilled checkpoint optimized for fast inference. Krea explicitly recommends training on Raw and running the resulting LoRA on Turbo.

Krea's hosted Turbo experience can generate in about two seconds, making it particularly useful for fast creative loops.

In our review of published workflows, one 80-image local comparison used Turbo at 8 steps and roughly 7 seconds per image, while a Raw plus Turbo-LoRA setup used 16 steps and roughly 22 seconds. The faster setup showed more repeated centered compositions, while the Raw-based workflow produced more seed and composition diversity. This is a workflow case, not a universal benchmark.

How Does Krea 2 Identity Edit Work?

Krea 2 Identity Edit Is a Community Extension

The Identity Edit v1.2 workflow covered here is not an official Krea.ai feature. It is a community fine-tune built on Krea 2 Raw for reference-conditioned editing.

Its role is fundamentally different from S-Ref: change clothing, pose, expression, props, or context while trying to preserve the person or subject.

The strongest prompting pattern is short and final-state focused. Instead of describing an editing operation, describe what the finished image should contain. Complex changes are more reliable when chained:

Original → Outfit → Pose → Expression → Prop

Identity Edit Settings That Matter Most

The current model documentation recommends Turbo at 8–12 steps with CFG 1.0 for most edits. Reference Boost around 4 is a strong-likeness starting point; values below 1 allow more freedom, while values above 10 can begin breaking removal tasks. Around 8 steps favors composition adherence, while 12 favors more facial detail.

For large removals, the documented starting point shifts to Raw at about 20 steps and CFG 3.0. The model author also recommends generating at no more than roughly 2MP to reduce duplication and source bleed. In two-input edits, input order is fixed: scene first, person second.

What Are the Real Limits of Krea 2 Identity Editing?

Identity preservation should be treated as a design objective, not a guarantee.

Across the model documentation and our review of user questions, the recurring failure patterns include facial-geometry drift, invented clothing details, non-local lighting or color changes, skin-tone shifts, identity convergence in multi-person scenes, and seed dependence. Outfit replacement is generally less predictable than a simple face-focused change.

One reviewed ComfyUI workflow ran on an RTX 3060 with 6GB VRAM and 16GB RAM, including a two-stage sampling setup and 2× upscale. The workflow was technically feasible, but outfit editing remained inconsistent. Hardware feasibility and production reliability are different questions.

For ecommerce, recurring characters, or advertising, every identity-sensitive output should therefore be compared against the source before approval.

Do You Still Need a Krea 2 LoRA for Character Consistency?

For occasional edits, Identity Edit may be sufficient. For a character that must recur across dozens of scenes, angles, and assets, a Character LoRA provides a learned identity prior instead of repeatedly inferring identity from one reference.

Krea's official training workflow requires at least three images, uses captions, and supports styles, characters, objects, and materials. Krea recommends Medium as the faster training starting point and Large when a stronger LoRA is needed.

In one character workflow from our research set, 40–50 images and roughly 1,000–1,500 steps produced strong perceived likeness on 16GB VRAM, although that likeness was subjective rather than a benchmark. Another style-LoRA case began overfitting at roughly 2,500–3,000 steps, making checkpoints every 250–300 steps useful for catching an earlier, cleaner model.

For video production, a related workflow created 16 pose and camera variations from one starting character before sending selected views into video generation. That reference-bank approach gives downstream models more structural information than relying on a single front-facing image. A similar principle appears in the MiniMax H3 reference guide.

What Is the Best Krea 2 Workflow for Each Design Goal?

For professional design production, the strongest sequence is Explore → Refine → Lock Style → Build the Aesthetic System → Control Identity or Structure → Edit One Variable at a Time → Build Variations → Finalize.

FAQ

What S-Ref strength should I use in Krea 2?

Start around 50% for a balanced relationship between prompt and reference. Around 20% keeps the prompt dominant, while 80% gives the reference much more control. Higher strength can improve style fidelity but also increases the risk that reference colors, composition, or subject details leak into the output.

Is Krea 2 Identity Edit official, and why can identity still drift?

No. The v1.2 Identity Edit workflow discussed here is a community LoRA built on Krea 2 Raw. Identity can still drift because edits regenerate more than a perfectly isolated local region. Facial geometry, clothing, skin tone, lighting, and unrelated details can change, especially when several transformations are requested at once.

Raw or Turbo: which Krea 2 model should I use?

Use Turbo for fast inference and iteration. Use Krea 2 Raw for fine-tuning, LoRA training, and advanced base-model workflows. Do not confuse the Raw model with Creativity Raw, which simply disables prompt expansion inside Krea's creativity control.

Do I need a LoRA for Krea 2 character consistency?

Not for every edit. If you only need a few changes to an existing image, an identity workflow may be enough. For recurring characters across many poses, scenes, angles, or video assets, a Character LoRA is more appropriate because identity is learned across a dataset rather than inferred repeatedly from one reference.

Conclusion

Krea 2 becomes much easier to control when styles, references, and identity are treated as separate design layers rather than problems for one perfect prompt to solve. Prompts define content, S-Refs transfer precise visual language, Moodboards establish a broader aesthetic universe, Generative Sliders shape exploration, and LoRA or identity-focused workflows handle recurring subjects. The most reliable production process is therefore not one-shot generation but exploration, style locking, identity or structure control, staged editing, variation building, and final production.