Note: This description was translated with AI assistance and may not perfectly capture my original intent. I plan to review it before uploading, but wanted to flag that in advance.

This LoRA is a more realistic reinterpretation of Erika, a character from The Secret Atelier, a Japanese visual novel. You can also find it on Steam.

The original character had an overly red color tone and a somewhat flat look, a byproduct of the game's particular 3D art style. To address this, I adjusted the color grading and applied several techniques to add realism, then used the original artwork as a reference to regenerate images, which became the training data. In other words, this is a character recreated with a realistic feel, and some drift from the original occurred in the process — though I tried to preserve as much of the original character's identity and personality as possible. I'm not sure this nuance will fully come across, but half-jokingly, I'd describe the result as less "Erika" and more "Serika."

If you simply prompt "woman," you'll get something closer to a Western person cosplaying as Erika. Using prompts like "young japanese woman" or "east asian woman" brings the output closer to the original Erika. Erika is a fictional character, but based on her design, I don't think she reads as strictly Japanese or East Asian — more mixed-heritage. In any case, feel free to generate whichever interpretation suits your taste. The showcase images were generated using "young japanese woman."

Training was done at rank/alpha 32, but the uploaded file has been reduced to rank 1. I generated roughly 100 comparison images between the two, and the difference is minimal. The rank 32 model does produce slightly stronger color saturation and a more pronounced character likeness, while the rank 1 model is a touch weaker on character likeness but reads as more photorealistic. Since I don't consider the difference significant, I'm only uploading the rank 1 file — it's just 8MB.

After generating more images later, I discovered a problem. I later realized that I had deleted more information than necessary because I was not familiar with the LoRa resizing method. I apologize to those who downloaded it first. Please download the new 69MB file again. It was foolish of me to think that it could be fully displayed within 8MB. Therefore, I am uploading the original 32-rank file, which was resized using a different method.

I won't be uploading explicit content, though I did generate some for testing purposes. Since the original character is already quite NSFW-coded, the LoRA produces strong NSFW results as well — feel free to try it yourself if you're curious.

The showcase images were generated using only this LoRA, with no additional LoRAs and no post-processing (first-pass output). The base model used was Krea2-Turbo-HD-V1-int8_convrot, which performs nearly identically to the official model. Using models like Gonzalomo or Moody will yield even more photorealistic results.

Training details (for anyone getting started with LoRA training):

Tool: musubi-tuner (kohya-ss)

Recently added --convrot_int8 option, which significantly reduces training time

Settings: dim 32 / alpha 32 / LR 0.0001 / LR scheduler: cosine / warmup 90 steps (5%) / total 1,800 steps / blocks to swap: 14

Dataset: 60 selected frames extracted from the source material, color-corrected, then upscaled and enhanced for realism using Qwen Image Edit 2511

Captioning: Generated with Qwen3.6 27B, then manually refined

Hardware: 32GB system RAM / RTX 4060 Ti 16GB

Total training time: ~6.5 hours