Character-LoRA-Anima

Overview

This research extends the Character-LoRA project based on the Anima model, focusing on research and testing with the BACKGWA character. The original SDXL-based model is Character-LoRA (BACKGWA), and the related research repository is BackGwa/Character-LoRA.

Please note that Anima differs from the original research in several aspects, including dataset labeling, training configuration, and other setup details. Therefore, this repository and model should be understood as a separate Anima-based research and testing project, rather than a direct continuation of the original SDXL-based model.

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.
• By using this LoRA, the user assumes full and exclusive responsibility for any legal disputes, social issues, infringement of third-party rights, defamation, or unlawful acts that may arise.
• Users must comply with all applicable laws, platform rules, and ethical standards in their country or region when using this LoRA. Any consequences resulting from violations of such laws, rules, or standards shall be borne solely by the user.
• Under no circumstances shall the creator, distributor, or maintainer of this LoRA be liable for any direct, indirect, incidental, special, or consequential damages arising from the use or inability to use this LoRA.
• Redistribution of this LoRA by third parties is prohibited in principle. However, it may be permitted if the original creator is clearly credited and proper attribution is provided.

Trigger Word

Use the trigger word below to activate the BACKGWA character. You can add extra prompts for outfit, expression, pose, background, or composition as needed.

backgwa

Usage

This LoRA was designed for use with Anima-based model environments. For the most stable and consistent results, it is recommended to use it with the anima-base-v1.0 model.

Preview

<Gallery />

---

Training

This model is not the result of a single standalone training run. It was produced through dataset reconstruction, multiple experimental training runs, LoRA/checkpoint merging, and additional fine-tuning. Therefore, the dataset size, epoch count, and training settings below are provided as references for the research process, and the final model metadata may not fully represent the entire training history.

The main training dataset was organized into three subsets:

The dataset was rebuilt using selected images from previous datasets based on quality and suitability, along with images generated using the Release 1 model. The overall style distribution was adjusted to reduce overfitting to a fixed visual style. Specific artist names were not directly used when creating the dataset, in order to avoid explicitly relying on a particular artist’s style.

During the research process, several dataset and training configurations were tested, including larger reconstructed datasets, stricter labeling, removing or rebalancing backgwa_additional, and splitting the dataset into separate subsets with adjusted repeat counts. These experiments produced partial improvements, but did not fully resolve style bias, prompt-response instability, or character reproduction issues.

For the final approach, multiple LoRAs and checkpoints were selected and merged. A total of 5 models were used in the merge, including Research Preview, Release 1, and two checkpoints from the Release 2 research process. The merged model showed better results than Research Preview, Release 1, and the intermediate experimental models. To reduce remaining issues inherited from Release 1, the merged LoRA was further trained using a dedicated 19-image dataset with a learning rate of 2e-5. This additional training was run for 20 epochs, and the 10th epoch result was used.

Research Repository

The research document and related materials are available at: BackGwa/Character-LoRA

License

The LoRA model is released under the circlestone-labs-non-commercial-license.
The research document and repository materials are released under the MIT License, unless otherwise specified.