Update README.md

README.md

@@ -34,14 +34,18 @@ license:
 ---
 
 # Overview
-Character-LoRA is an SDXL-based character LoRA created to reproduce the `BACKGWA` character. This LoRA was trained on a copyright-conscious synthetic image dataset.
 
-This project was conducted for research purposes to protect the copyright, creative style, training rights, and intellectual property held by the original artist of this character.
-Instead of using images directly created by the artist as training data, this LoRA was trained only on an AI-generated synthetic image dataset.
+Character-LoRA is an SDXL-based character LoRA created to reproduce the `BACKGWA` character using a copyright-conscious synthetic image dataset.
 
-Through this approach, the project explores a way to build a character LoRA without directly using the original character images or the artist's work as training data.
+This LoRA was developed as part of research on whether a character LoRA can be trained without directly using original character images or artist-created works as training data.  
+Instead of collecting or reusing human-made artwork, the training dataset was constructed from AI-generated synthetic images based on structured character descriptions.
+
+The full research document, methodology, instruction template, and JSON schema are available in the GitHub repository: [`BackGwa/Character-LoRA`](https://github.com/BackGwa/Character-LoRA)
+
+Through this approach, the research examines a workflow for constructing character LoRA models while reducing dependence on original images and considering copyright, training rights, artistic style, and intellectual property.
 
 ## Disclaimer
+
 - This LoRA may be used freely for the user's own creative purposes. However, all copyright issues, legal responsibilities, and social or ethical consequences arising from any images, videos, or other outputs generated using this LoRA remain solely with the user.
 - Users must ensure that generated outputs do not infringe upon the rights, reputation, or creative works of the original artist, character rights holder, or any third party.
 - The creation and use of SFW and NSFW outputs are entirely at the user's own discretion and responsibility. The creator, distributor, and maintainer of this LoRA bear no responsibility or obligation regarding such outputs.
@@ -60,6 +64,7 @@ backgwa
 ```
 
 ## Usage
+
 This LoRA was designed for use with SDXL-based model environments.  
 For the most stable and consistent results, it is recommended to use it with the [`LUMIERE-Q`](https://huggingface.co/BackGwa/LUMIERE-Q) model, though it may also be used with other SDXL-compatible checkpoints.
 
@@ -71,26 +76,34 @@ For the most stable and consistent results, it is recommended to use it with the
 
 # Training
 
-This LoRA was trained without using any human-made artwork as training data.
+This LoRA was trained without using artist-created images or human-made artwork as training data.
+
+The dataset was generated with GPT Image 2.  
+A total of 80 synthetic images were generated as training candidates, and 48 images were selected for the final dataset based on consistency and quality.
+
+To prepare the character description used for image generation, the original character image and supplementary information were provided to the locally executed [`gemma-4-E4B-it`](https://huggingface.co/google/gemma-4-E4B-it) model.  
+The analysis used predefined instructions and a structured JSON schema to describe the character in a form suitable for subsequent synthetic image generation.
+
+The selected synthetic images were then labeled using GPT-5.5.  
+The labels describe visible attributes such as expression, pose, composition, and other image-specific details used during training.
 
-The dataset consists of 48 synthetic images generated with GPT Image 2.  
-These images were created specifically for this LoRA training process, and were not collected from, derived from, or reused from artwork created by human artists.
+The LoRA was trained using `sd-scripts` on the SDXL-based [`LUMIERE-Q`](https://huggingface.co/BackGwa/LUMIERE-Q) model.
 
-To prepare the initial image-generation prompts, the original character image was provided as a reference image to the locally executed [`gemma-4-E4B-it`](https://huggingface.co/google/gemma-4-E4B-it) model.  
-Predefined prompts and instructions were used to guide the model in generating text prompts capable of reconstructing the character from the provided reference image.
+| Parameter | Setting |
+| --- | --- |
+| Base Model | [`LUMIERE-Q`](https://huggingface.co/BackGwa/LUMIERE-Q) |
+| Training Tool | `sd-scripts` |
+| Dataset Size | `48 images` |
+| Generated Candidates | `80 images` |
+| Epochs | `10` |
+| Repeats | `10` |
+| Resolution | `1024x1024` |
 
-Using these reconstructed prompts, synthetic character images were generated with GPT Image 2.
+## Research Repository
 
-The generated images were then separately labeled using GPT-5.5.  
-These labels describe the visible visual features, outfit elements, composition, and other relevant details used during training.
+The research document and related materials are available at: [`BackGwa/Character-LoRA`](https://github.com/BackGwa/Character-LoRA)
 
-A total of 48 synthetic images were used for training.  
-The LoRA was trained on the SDXL-based [`LUMIERE-Q`](https://huggingface.co/BackGwa/LUMIERE-Q) model at a resolution of `1024x1024`, with 10 repeats and 10 epochs.
+## License
 
-|Parameter|Setting|
-|---|---|
-|Base Model|[`LUMIERE-Q`](https://huggingface.co/BackGwa/LUMIERE-Q)|
-|Dataset Size|`48 images`|
-|Epochs|`10`|
-|Repeats|`10`|
-|Resolution|`1024x1024`|
+The LoRA model is released under the CreativeML Open RAIL-M license.  
+The research document and repository materials are released under the MIT License, unless otherwise specified.