Krea2 Character LoRA Collection

A collection of 84 character LoRAs for the KREA2 base AND turbo model, each paired with a persona .md file describing the character's stable facial features, variable appearance, and expressions.

Every LoRA ships as <Name>_KREA2.safetensors alongside <Name>_KREA2.md inside its own folder under Characters/.

Updates

Repository restructure

• All characters moved under a single Characters/ directory (one folder per character containing the .safetensors, the persona .md, and the training captions).
• A new Pruned_Comparison folder was added with one preview image per character (see below).

Dynamic re-rank + block pruning

Every LoRA was re-ranked with dynamic sv_fro SVD (retention 0.998, max rank 16) and block-pruned by dropping the 8 lowest-impact transformer blocks (blocks 1-8, ~3% of delta energy).

Retrained character

• Jung Ho-yeon (Ho-yeon) was retrained on a new dataset, then re-ranked and pruned like the others.

Fully automated pipeline

This repository - dataset, training, and upload - was produced end-to-end by an autonomous Hermes Agent (served by the Kimi K3 model) from a natural-language request. No manual review gates anywhere in the loop:

1. Scraping - automated image collection per character from web photo pools.
2. Cropping & curation - face/head detection Ultralytics Yolo and crop, with a hard content filter (No AI generated images, correct subject only, sharpness threshold).
3. Aesthetic gating - every candidate image scored by a discriminator built on the ERNIE aesthetic image scorer; low scorers rejected automatically with a hard pass/fail threshold.
4. Captioning - datasets captioned with Gemma-4 26B (abliterated) vision MTP for unrestricted, descriptive captions, fast (1.2s/i).
5. Persona synthesis - each <Name>_KREA2.md generated from caption trait-analysis (facial features, range of expressions).
6. Infrastructure - training pods created, configured, monitored, and torn down automatically via the RunPod API (MCP).
7. Training - LoRAs trained on pods with Ostris AI-Toolkit with user predefined settings.
8. Publishing - weights and persona files uploaded to this repo by the agent.

Pruned comparison

One example image per character, generated with the pruned LoRA. Click a character to expand.

Known limitations

• Trained for likeness; style flexibility depends on your prompt and sampler settings.

Disclaimer

These LoRAs reproduce the likeness of real people and are shared for research and creative purposes. Please use them responsibly and respect the rights of the individuals depicted - don't create misleading, harmful, or non-consensual content. By downloading, you take responsibility for what you generate.

Credits

• Pipeline automation: Hermes Agent (Nous Research) on Kimi K3
• Training: Ostris AI-Toolkit
• Aesthetic scoring: ERNIE aesthetic scorer (CRT-Nodes ComfyUI packaging)
• Captioning: Gemma-4 26B (abliterated) vision MTP
• Infrastructure: RunPod
• User Discord https://discord.gg/uV6ZNgg33j