About Rina v2

Normally, when a program gets a version update, you would expect improvements in performance or new features. However, this is not a program, and I cannot guarantee that this version is necessarily better than the previous one. Please think of it simply as a different version.

The differences between v1 and v2 can be seen in the showcase images.

What is different about v2?

The dataset for my original Rina v1 LoRA was made from screenshots of the original game. Many of those images are quite explicit and contain a lot of nudity. That is simply the nature of the game.

As a result, when the LoRA was applied, it would often produce unintended sensitive nudity even when it was not specified in the prompt. This was especially noticeable when using models without safety filters.

v2 was created to address this problem of unintended nudity.

At the same time, I wanted to make Rina look more like a realistic person. During this process, my own preferences and interpretation became part of the result. Because of that, v2 is arguably closer to "Mina" or "Nana" than "Rina." I consider her to be a derivative/fan-created interpretation of the character rather than a faithful recreation of the original.

The v2 dataset was created from images generated using the v1 LoRA. Unlike v1, v2 can produce a more realistic-looking person without having to stack a large number of LoRAs just to increase realism.

However, since v2 itself was created using v1, there is little reason to download v2 if you already know how to use v1 to create a convincing realistic version of a fictional character.

Several intermediate versions were created while changing the settings and dataset along the way, but they all had the problem of increasingly strong AI-generated image artifacts and characteristics. A lot of trial and error went into trying to solve this.

Although I call it v2, internally it is probably closer to something like v8.

I also tried adding regularization images, but they did not help very much. To be honest, I am not completely satisfied with the v2 model I am uploading now either. However, I no longer know what else I could do to improve it, and I simply do not have enough time to run more experiments.

My potato PC takes way too long to train these models.

I intentionally do not remove the metadata. If you are curious about the training settings and how this LoRA was trained, feel free to check the metadata.

v1 LoRA Description

This is a recreation of a character from a minor Japanese game whose title cannot be disclosed.

Training was done at dim 32 / alpha 32, and the uploaded file has been size-reduced to rank 1. Testing showed that a high rank isn't necessary — rank 1 is sufficient to reproduce the character accurately. As a result, the file is only about 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 60MB 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.

If you want a more realistic rendering of the character, you can increase realism by using loraholic's "Real vs 3D" slider LoRA, or the Krea2 Edit "Anime to Real" LoRA. Another option is to use a different base model that has been fine-tuned for realism.

The showcase images were generated at 832x1216 resolution, combining several methods in a single first pass. This means the same results cannot be reproduced using this LoRA alone. The original character is a 3D render, and applying this LoRA by itself will produce a flat, plain 3D-style character similar to the source material. Lowering the LoRA strength to around 0.5 reduces character fidelity but gives a more human-like appearance.

Training Specs and Settings

The reason I'm including this section is that when I first started training LoRAs, the hardest part was finding other people's documented experiences to reference. So I'm leaving this record for others who are just getting into LoRA training, in hopes that it helps.

Training tool: kohya-ss's musubi-tuner

Training settings:

• dim 32 / alpha 32 / LR 0.0002 / LR scheduler: cosine / warmup: 90 steps (5%) / total steps: 1,800

• blocks to swap: 24

Dataset: Selected 60 images out of roughly 150 screenshots captured from the game, upscaled using Qwen Image Edit 2511. Captions were auto-generated with Qwen3 VL 8B and then manually corrected where inaccurate.

System specs: 32GB system RAM / RTX 4060 Ti 16GB

Total training time: approximately 9 hours