ANDREAV2
A fully custom AI character trained from scratch — sculpted in 3D, refined through AI realism, and distilled into 28 carefully curated images using the ostris ai-toolkit.
Building a person from nothing
AndreaV2 didn’t begin with photos. She began with vertices. The likeness was sculpted entirely in 3D, giving precise control over facial structure and proportions before a single photorealistic pixel existed.
From those 3D renders, Google Gemini’s image generation was used to breathe photorealism into the base — transforming clean 3D geometry into convincing skin, light, and depth. The results were then hand-curated down to 28 prime training images, each chosen to maximize the LoRA’s expressiveness and consistency across styles.
Training was run locally using the ostris ai-toolkit, with Claude Code guiding configuration, parameter tuning, and troubleshooting throughout the process.
Three Phases. One Character.
The foundation. Andrea’s facial structure, proportions, and likeness were built entirely inside a 3D modeling program — giving complete ownership over every geometric detail before any AI was involved.
Multiple lighting setups and camera angles were rendered from the 3D base to establish variety in the training data from a single, consistent source.
3D renders were passed into Google Gemini’s image generation to layer photorealistic skin, hair, and lighting over the clean geometric base.
This step transformed synthetic renders into images that read as real photography — creating a training set that was both consistent in identity and diverse in photographic style.
The curated 28-image dataset was fed into ostris ai-toolkit running fully local — no cloud, no data sharing. Captions were crafted per-image for maximum trigger consistency.
Claude Code assisted with toolkit configuration, training script setup, and live troubleshooting — turning what’s typically a painful setup into a streamlined, repeatable pipeline.
Training Data & Results
Exceptional Diversity in Outputs
Sample Selection from the Training Dataset
3D Renders Serve as Starting Point
AndreaV2 in Motion
Video generations using the AndreaV2 LoRA across different workflows and animation pipelines.
Training Specifications
Full configuration details for reproducibility and transparency. Edit these values to match your exact settings.
WHAT’S NEXT
AndreaV3 is in development. Subscribe or check back for new LoRAs, workflows, and tutorials.