← sources

Sharing Asian Character LoRA for Krea 2(AI-Toolkit Training Settings Included)

source: reddit (reddit:StableDiffusion)published 2026-09-25 07:55ingested 2026-09-26 02:37

retained text

https://preview.redd.it/pvetvjp4ynrh1.jpg?width=814&format=pjpg&auto=webp&s=baf368a204d190bd1c14bcf822929f5ded9572d0

Hey everyone! Sharing my Asian Character LoRA trained specifically on **Krea 2 Raw and Turbo**.

You can check out the showcase and download it out here:

👉 [**Hugging Face Space Showcase**](https://huggingface.co/spaces/fidel1234/Krea2_Asian_Character_Showcase?utm_source=gemini)

# 📷 Dataset & Framing

* **Hand-Cherry-Picked Dataset:** I usually use photobook as the dataset, and combination of qwen, gemini, and grok to caption it. I usually use around 20-30 images, quality over quantity
* **Balanced Framing:** Includes a varied mix of close-up, medium shot, and full-body shot angles

**Note:** Sometimes I use image editor (photoshop/flux klein/gemini) to edit and seedvr to upscale. Creating lora with old photo is really hard! Krea is sensitives with moles, character that has mole around her face usually will destroy the output, I recommend to remove them instead

# ⚙️ Training Parameters (AI-Toolkit)

For those interested in the technical setup or looking to replicate/tweak the workflow, here are the full settings from AI-Toolkit:

* **Base Model:** Krea 2 (raw) (`krea/Krea-2-Raw`)
* **Trigger Word:** `xxxx`
* **Linear Rank:** 32
* **Quantization / Compile:** `qfloat8` (Transformer & Text Encoder)
* **Save Data Type:** BF16
* **Optimizer:** AdamW8Bit
* **Learning Rate:** 0.0001 | **Weight Decay:** 0.0001
* **Steps:** 2500 | **Batch Size:** 1 | **Gradient Accumulation:** 1
* **Timestep Type / Bias:** Sigmoid / Balanced
* **Loss Type:** Mean Squared Error
* **Text Encoder Optimizations:** Unload TE (Disabled), Cache Text Embeddings (Enabled)
* **Resolutions Multi-Aspect Bucket:** 512, 768, 1024
* **Caption Dropout Rate:** 0.05 (`.txt` extension)

It's trained on RTX 5090, usually take around 1 hour 15 mins for each lora

I will add more in future! It's becoming one of my hobby now so.

Happy Friday, and have a nice weekend.

*Disclaimer: Please use responsibly and respect the rights of depicted individuals—do not create misleading, harmful, or non-consensual content. By downloading, you take responsibility for what you generate.*

document fetched 2026-09-26 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1wpqd82/sharing_asian_character_lora_for_krea_2aitoolkit