Krea 2 Bianca V7 — Character LoRA

# Krea 2 Bianca V7 — Character LoRA

Bianca V7 is a photorealistic character LoRA for Krea 2, representing the

fictional adult character Bianca. It was trained on Krea 2 Raw and

validated in ComfyUI on Krea 2 Turbo.

The goal was never just to reproduce one attractive face in one familiar

portrait. We wanted a character who could remain recognizably Bianca while

acting, reacting and living inside a scene: laughing naturally, looking away

from the camera, moving, working with her hands, wearing detailed clothes,

appearing in difficult light and occupying environments that were not present

in the training set.

## A human + AI collaboration

This LoRA was developed as a long-form collaboration between a human, acting as creative director and final human evaluator, and ChatGPT/Codex, acting as a technical research and production partner.

The human role was decisive throughout the project: defining who Bianca is,

judging whether a result truly looked and felt like her, identifying unwanted

age drift or over-stylization, raising the difficulty of the tests and making

the final approval call. The AI-assisted role covered dataset auditing,

caption and manifest engineering, reproducible configuration, ComfyUI

automation, training supervision, checkpoint preservation, hash verification,

controlled comparison and technical documentation.

That division mattered. Identity similarity scores can help organize a test,

but they cannot decide whether an expression feels alive, whether the face has

aged incorrectly or whether a technically clean image still misses the

character. Bianca V7 was therefore approved by human visual judgment supported

by measurements, not replaced by them.

## From V6 to V7

V7 was not made by simply training V6 for more steps. We began with an

individual audit of 55 image-caption pairs and preserved the material that

was already useful. Weak, redundant or misleading examples were not allowed

to stay merely to increase the dataset size.

The final V7 dataset contains 53 deduplicated image-caption pairs:

- 41 preserved and corrected examples from the audited material

- 12 purpose-built additions for specific gaps

- captions reviewed for identity, emotion, gaze, pose, action, framing,

lighting, clothing and environment

- manifests and hashes connecting every training item to its provenance

The corrective images were selected to teach missing behavior rather than

repeat easy beauty portraits. The aim was coverage: spontaneous emotion,

off-camera gaze, movement, practical action, profiles and three-quarter views,

hands interacting with objects, full-body framing, difficult lighting,

layered clothing and visually busy locations.

## What did not work — and what we learned

This release also contains the lessons from failed and weaker attempts:

- Some early prompts made Bianca look older than intended even when the image

was otherwise attractive.

- Low LoRA strengths preserved scene freedom but weakened identity too much.

- High identity pressure could flatten expression or make the result feel more

like a posed reference photo than a living scene.

- More training was not automatically better: the step-2000 checkpoint was

complete and valid, but step 1500 preserved a better balance of identity,

naturalness and scene plasticity.

- A controlled V6 img2img stress test produced black outputs on a few

difficult references. The failures were reproduced, documented and resolved

with safer conditioning instead of being silently removed.

- Strong img2img similarity can be useful for stress testing, but it was kept

separate from the frozen text-to-image validation used to approve the LoRA.

These findings changed the dataset, the prompt protocol and the final

checkpoint decision. The released file is the checkpoint that performed best,

not merely the last file produced by the trainer.

## Recommended settings

- Trigger: BiancaV7

- Recommended strength: 0.70–0.75

- Balanced starting point: 0.70

- Practical default after extended stress testing: 0.75

- Validated inference base: Krea 2 Turbo

- Sampler / scheduler: Euler / Simple

- Steps: 8

- CFG: 1.0

Below 0.50 the identity becomes substantially weaker. The selected checkpoint

is step 1500. Strength 0.70 passed the frozen formal protocol; 0.75 was then

preferred by the creative director across a broader range of ordinary,

cinematic and extreme stress-test scenes.

## Reproducible training

- 53 audited image-caption pairs

- 1024 logical resolution with buckets

- rank / alpha: 32 / 32

- AdamW8bit, constant LR 1e-4

- bf16, FP8-scaled base, SDPA, gradient checkpointing

- seed 42

- A40 48 GB

- selected checkpoint: step 1500 of 2000

- checkpoints preserved every 250 steps

- smoke test completed before the production run

## Controlled validation

Prompts, seeds, dimensions, sampler settings and workflows were frozen so each

comparison changed only the intended model or strength. The release passed:

- 48-image checkpoint screening

- 36-image strength sweep

- 18-image unseen holdout against Bianca V6 and pure Krea 2

- six unseen scenarios covering emotion, gaze, action, environment and light

- no hard failures in the final holdout

The final holdout averages were:

- identity consistency: 3.08 / 5

- global image quality: 3.67 / 5

- expression and gaze: 3.83 / 5

- anatomy and action: 3.83 / 5

- scene plasticity: 4.00 / 5

After formal approval, Bianca V7 was pushed further through ordinary life,

fashion, low-key light, neon, crowded interiors, action, difficult camera

angles and reference-guided img2img stress tests. Those additional tests did

not replace the holdout; they demonstrated how the approved LoRA behaved

outside it.

The accompanying gallery uses original text-to-image prompts rather than

third-party image references. Workflows, scoring sheets, contact sheets,

captions, manifests, logs, costs and SHA-256 hashes were retained as part of

the project record.

## Why we are publishing it

Bianca V7 is more than a downloadable weight file. It is the result of a

traceable creative experiment: preserve what works, document what fails,

change one variable at a time and let a human creative judgment make the final

decision. We are publishing both the character and the process because

reproducibility should not remove personality, and personality should not

require abandoning technical rigor.

## Limitations

Identity can weaken at extreme distance, behind heavy occlusion, with very

unusual expressions or below the recommended strength. Hands, typography and

small objects remain subject to limitations of the base model.

Bianca is a fictional adult. Do not use this model for fraud, deceptive

representation of real people, harassment, illegal content or violations of

privacy, consent or intellectual-property rights.

## License and attribution

Krea 2 is licensed under the Krea 2 Community License Agreement:

https://www.krea.ai/krea-2-licensing

This LoRA is a modified derivative of Krea 2 and is not an official Krea

product or endorsed by Krea. Commercial use is subject to the revenue threshold

and all other conditions in the Krea 2 Community License Agreement.

## Nota em português

Bianca V7 nasceu de uma parceria real entre direção criativa humana e execução

técnica assistida por IA. Nós auditamos o material imagem por imagem, mantivemos

o que funcionava, corrigimos lacunas específicas, treinamos com checkpoints

rastreáveis e recusamos escolher o último checkpoint apenas por ele ser o

último. A aprovação final veio da combinação entre testes congelados,

documentação completa e avaliação visual humana. O resultado é uma Bianca mais

expressiva, flexível e capaz de continuar sendo ela mesma em situações fáceis,

difíceis e extremas.