Premium character LoRAs

Original synthetic characters, bundles and AI tools are in my shop:

https://rkrra.com/civitai

Get the character LoRAs on Patreon: https://www.patreon.com/cw/RkRRa

Join the community on Discord: https://discord.gg/nB79wRNtMT

Kate Raka

A character LoRA for Krea 2, built entirely from synthetic images. No photographs of real people were used at any stage.

Usage

Include the trigger word kate_raka in your prompt.

Recommended weight: 1.

How this was made

The character began as 3 reference images. Those showed her from the waist up, which meant nothing existed to describe her legs, back, or full proportions -- so a two-stage process was used rather than training directly on them.

Stage one -- the pose set. A written description of her fixed physical traits was produced, then used to generate 8 views under identical conditions: full body from the front, three-quarter, profile and behind; half body from two angles; and head-and-shoulders portraits. Same wardrobe, same neutral studio lighting, same plain background in every one. Those were trained into a small bootstrap LoRA at rank 32 over 1500 steps.

Stage two -- the training set. That bootstrap generated 110 images across a deliberate spread of framing, camera angle, expression, lighting and setting, weighted toward full and half-body shots. Each was reviewed by hand and a large share discarded -- for identity drift, anatomical errors, or simply being redundant. The survivors, plus the original pose set, form the final dataset.

Why this many images

118 images. A character LoRA needs enough variety to separate the person from the circumstances they appear in, and not so much that it simply memorises them. Below roughly forty, the model tends to bind incidental details -- a background, a light direction, a pose -- into the character itself. Past a hundred, additional images mostly repeat what earlier ones already taught.

The composition matters more than the count. The 8 studio poses anchor the body proportions under controlled conditions; the generated images teach that the same person holds up across different clothing, lighting and settings. Neither alone would be sufficient.

The images were reviewed against the original references rather than against each other -- comparing generations to other generations normalises toward whatever drift they share.

Training

• Base model: Krea 2 (trained on Raw)
• Type: Character LoRA
• Dataset: 118 images (110 generated, 8 pose set)
• Steps: 2500
• Network rank / alpha: 32 / 32
• Learning rate: 0.0002
• Optimizer: adamw8bit
• Resolutions: 768 / 1024

Generation settings

Trained on Krea 2 Raw, and transfers to Krea 2 Turbo for faster generation.

• Raw: ~30 steps, guidance 3.5
• Turbo: ~8 steps, guidance 1.0

The preview images were generated with krea2_turbo_mxfp8.safetensors, at 8 steps, guidance 1, LoRA weight 1.