<h1 id="overview">Overview</h1>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 <a target="_blank" rel="ugc" href="https://civarchive.com/models/2600891/character-lora-backgwa">Character-LoRA (BACKGWA)</a>, and the related research repository is <a target="_blank" rel="ugc" href="https://github.com/BackGwa/Character-LoRA">BackGwa/Character-LoRA</a>. <h2 id="disclaimer">Disclaimer</h2><ul><li>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.</li><li>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.</li><li>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.</li><li>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.</li><li>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.</li><li>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.</li><li>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.</li></ul><h2 id="trigger-word">Trigger Word</h2>Use the trigger word below to activate the <code>BACKGWA</code> character.You can add extra prompts for outfit, expression, pose, background, or composition as needed.<pre><code>backgwa</code></pre><h2 id="usage">Usage</h2>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 <a rel="ugc" href="https://civarchive.com/models/2458426/anima">anima-base-v1.0</a> model.<hr /><h2 id="research-repository">Research Repository</h2>The research document and related materials are available at: <a rel="ugc" href="https://github.com/BackGwa/Character-LoRA">BACKGWA/Character-LoRA</a><h2 id="license">License</h2>The LoRA model is released under the <code>circlestone-labs-non-commercial-license</code>.The research document and repository materials are released under the MIT License, unless otherwise specified.

BACKGWA-Anima

BACKGWA Anima Model Release 2 has been released.This release improves the reproduction stability, prompt responsiveness, composition diversity, expressive range, and labeling quality of the BACKGWA character. It focuses on reducing issues observed in previous versions, including excessive fixation on specific visual styles, limited expressions and compositions, partial character design errors, and unintended yellow elements appearing too frequently.This model is not the result of a single standalone training run. It was produced by merging multiple LoRAs and checkpoints, followed by additional training. Therefore, the dataset sizes, epoch counts, and training settings described below are provided as references for the research process, and the final model metadata may not fully represent the entire training history.Training Dataset Improvements<ul><li>The new dataset was built from selected images from previous datasets based on quality and suitability, along with images generated using the Release 1 model.</li><li>To reduce the issue where the image style became overly fixed in the previous version, the overall style distribution of the dataset was made more diverse.</li><li>Specific artist names were not directly used, and visual diversity was instead achieved through minimal style prompting techniques. This was done to avoid creating training data that explicitly relies on a specific artist’s style.</li><li>Missing or incorrect labels were corrected, and the model was improved to follow natural language prompts more reliably.</li><li>Composition diversity, expressive range, and character reproduction stability were improved, while some character design errors and unintended yellow color artifacts were partially reduced.</li></ul>Datasets Used<ul><li><code>backgwa_base</code>: A dataset for learning the basic appearance and core features of the character. It contains 40 images and was repeated 2 times per epoch.</li><li><code>backgwa_alignment</code>: A dataset for preserving the character’s invariant appearance and identity while reducing bias introduced by <code>backgwa_additional</code> and biases observed in Release 1. It contains 20 images and was repeated 4 times per epoch.</li><li><code>backgwa_additional</code>: A dataset for improving expressiveness and prompt responsiveness under various transformation conditions, including body type, age, clothing coverage variations, and detailed visual depiction. It contains 40 images and was used 1 time per epoch.</li></ul>Training Settings<ul><li>The initial training was conducted for 40 epochs with a learning rate of 0.00001, and the 40th epoch result was used for this stage.</li><li>Afterward, additional training was performed on the merged LoRA using a dedicated small-scale dataset of 19 images.</li><li>This additional training was conducted for 20 epochs with a learning rate of 0.00002, and the 10th epoch result was used.</li></ul>Research and Experimentation Process<ul><li>In the Research Preview stage, the basic character design was reproduced relatively well, but the model tended to generate repetitive images with similar compositions, expressions, and styles.</li><li>The keywords, prompt structure, and training parameters were then revised, which produced some improvements. However, new issues appeared, such as exaggerated body proportions and unstable clothing or color selection.</li><li>Negative prompts and more precise prompting could reduce some issues, but they were not sufficient to resolve the model’s internal bias.</li><li>Several experiments were conducted, including reconstructed datasets of 150 and 100 images, stricter labeling for <code>backgwa_additional</code>, removing <code>backgwa_additional</code>, splitting the dataset into separate subsets with adjusted repeat counts, and rebalancing dataset weights.</li><li>Each experiment produced partial improvements, but the final results were still not satisfactory due to issues such as style bias, loss of detailed expressiveness, and excessive influence from the additional dataset.</li></ul>Final Approach<ul><li>Since retraining with a single dataset was not sufficient, existing LoRAs and checkpoints from the research process were selected and merged.</li><li>A total of 5 models were used in the merge, including Research Preview, Release 1, and two checkpoints from the Release 2 research process.</li><li>The merge candidates were selected based on relatively low bias, stable character reproduction, and stable image quality.</li><li>The merged model showed better results than Research Preview, Release 1, and the intermediate experimental models.</li><li>However, some issues from Release 1 remained, so additional training was performed using a dedicated small-scale dataset of 19 images.</li></ul>NotesThis release should be understood as a research-stage improvement for increasing BACKGWA character reproduction stability and prompt responsiveness, rather than a final result that fully resolves all previous issues. Because this model was created through both merging and additional training, its metadata may not fully reflect the entire training process.

Files with SHA256: 22d2e63f62e36a14c8c4f21d2d325d3b23799a95f357fda194c9826c140b97ea