Krea 2 Valentina

Made for DearAI: https://dearai.live/studio/valentina.

Valentina is a fictional 29-year-old Afro-Brazilian character: expressive curls, a warm smile, and a recognizable face across portraits, everyday scenes and different outfits. This first release includes quiet kitchen moments, city evenings, dancing and a walk by the sea.

If you enjoy using her, an honest review and your own sample images help other creators judge the model.

Start here

- Trigger: zxvalentina woman, 29 years old,

- Base: Krea 2. Start with Krea 2 Turbo.

- LoRA strength: 1.0, the tested setting.

- Turbo settings: 8 steps, CFG 1, Euler / Simple, 1024 × 1344.

- Name the outfit, action, lighting and setting you want. Unload the LoRA when generating another character; removing the trigger alone does not reliably remove her identity.

Example:

zxvalentina woman, 29 years old, waist-up, 35mm film photo, mid-turn glancing back over her shoulder, serious expression, hair loose, walking under the Lapa arches at night, a denim jacket over a white top, on-camera flash, cool night grain, 35mm, photograph

This is the cover prompt. The result suggests a night-time archway; it is not a geographically exact reconstruction of Lapa.

About the sample images

The eight launch samples use Krea 2 Turbo fp8 + Valentina at 1.0 + Realistic Snapshot Krea2 at 1.0. The additional realism LoRA is not included in this download. They use an initial 1024 × 1344 render, a 1.5× Lanczos upscale and a three-step refinement at denoise 0.25, producing 1536 × 2016 images. These are refined outputs, not native renders at that final size.

Realistic Snapshot Krea2 resource: https://civarchive.com/models/2268008/realistic-snapshot-z-image-turbo-krea-2?modelVersionId=3084537. The PNGs retain their generation settings; SAMPLE_SETTINGS.json lists the exact prompts and seeds. The gallery's rendering style should not be attributed to Valentina alone.

Tested results

The selected checkpoint is step 1750. Across 16 character-prompt test images, the recorded mean face similarity was 0.802, with all 16 above the project's 0.60 threshold. Full-body and profile slot means were 0.704 and 0.731. These are face-similarity measurements, not percentages of realism or guarantees for new prompts.

A separate fixed-scene check at strength 1.0 gave face scores of 0.830 on Turbo, 0.782 on Sick Ollie and 0.795 on CyberRealistic v3. This is a limited compatibility check, not a broad benchmark of every scene or model version.

Known limitations

- Body proportions are not tightly locked. The fixed-scene outputs were slimmer than the reference build on all three tested bases. Descriptive body wording changed the result but did not fully recover the measured reference range.

- Nose jewellery and small facial marks can change, duplicate or disappear. Exact placement is not guaranteed.

- The tested showcase setup sometimes exaggerates sweat, wet skin and surface texture. The contributions of the character LoRA, the additional realism LoRA and refinement have not been isolated.

- A phone can appear when the prompt merely asks for a phone photograph. Props, mirrors, readable text and background people still need visual review.

- Subtle patterned texture can appear on plain backgrounds, particularly when contrast is increased.

- Wide shots have less facial detail than close portraits, and performance outside the tested settings is unmeasured.

License and attribution

This adapter is a derivative of Krea 2 under the Krea 2 Community License Agreement. A copy of that agreement and the attribution notice accompany the download. Use is subject to its terms and incorporated Acceptable Use Policy. This is not an official Krea product and is not endorsed by Krea.

Official Krea 2 license: https://www.krea.ai/krea-2-licensing.