Krea 2 — H01 Living Color

My first style LoRA for readable subjects inside large, living fields of color.

H01 turns people, animals, objects, interiors, and places into matte editorial paintings where the subject stays clear while broad organic color shapes become an equal part of the composition. Expect simplified silhouettes, restrained dark edges, dry-brush gaps, coarse paper texture, and a saturated earthy palette of coral, mustard, deep blue, black, cream, and muted green.

This is a style LoRA. It does not teach a fixed character, pose, costume, or recurring setting. The training set deliberately mixes figures, groups, animals, still lifes, transport, architecture, empty environments, and restrained fantasy so the visual language can travel between subjects.

Quick start

- Trigger: h01livingcolor

- Recommended starting strength: 0.65

- Useful exploration range: 0.45–0.85

- High-style range: 0.90–2.00 — stronger color fields and simplification; prompt fidelity may decrease

The trigger carries the shared material and color language. You usually do not need to repeat “gouache,” “textured paper,” or the palette. Add those words only when you want to push the effect harder.

What it does well

- readable figures and gestures inside simplified environments;

- bold editorial compositions built from a few large shapes;

- animals, still lifes, architecture, transport, and quiet fantasy;

- wide shots with a small subject and strong negative space;

- cloth, water, roads, shadows, foliage, and walls that continue into surrounding color fields;

- square, portrait, landscape, and vertical compositions.

Prompting notes

Describe who or what is present, what is happening, the spatial relationship, and the camera distance. Clear nouns and verbs work better than long quality-tag chains.

At lower strength, Krea 2 keeps more of its native modeling and detail. Around 0.65, the image usually has a stronger matte color-field construction. Above 0.85, silhouettes, palette, and surface can dominate the requested details. Treat these values as starting points and adjust for your loader and workflow.

Known limitations

- very high LoRA strength can flatten faces, hands, small props, and depth;

- detailed photorealistic or glossy prompts fight the learned material language;

- long prompts with many competing style words can weaken composition;

- complex multi-person contact and exact anatomy still depend heavily on the base model and seed;

- the model favors broad interlocking color regions and may simplify busy backgrounds.

Training information

- Base used for training: krea/Krea-2-Raw

- Intended inference target: krea/Krea-2-Turbo (also usable with Raw)

- Version: v1.0, rank 32, alpha 32, step 1200

- Resolution: 1024 with aspect-ratio buckets

- Training images: 60

- Trigger in every caption: h01livingcolor

- Captioning: short natural-language content and spatial descriptions; the common style description was intentionally omitted from captions so the trigger carries it

- Training precision: FP16

- Training-time transformer quantization: NF4 with double quantization

- Learning rate: 0.0003, constant schedule, no warmup

- Seed: 421

- Target modules: to_q, to_k, to_v, to_out.0, to_gate

Dataset and provenance

The 60-image training set was created specifically for this project with OpenAI ImageGen and manually curated from 66 candidates. The final set contains 22 people scenes, 13 animal scenes, 11 object scenes, 7 environments, 5 transport scenes, and 2 restrained fantasy scenes.

License and attribution

This LoRA is a derivative of Krea 2 and is distributed under the Krea 2 Community License Agreement v1 (June 22, 2026). Krea 2 was modified to create this LoRA. This is an independent community release and is not an official Krea product or endorsed by Krea.

Commercial use is subject to the Krea 2 license, including its revenue threshold and deployment requirements.

Required notice: “Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.”

Official license: https://github.com/krea-ai/krea-2/blob/main/docs/KREA-2-COMMUNITY-LICENSE

Please share your results and include generation metadata. It makes the model much easier for other people to learn.