⚠️ Model Access Restricted: Compliance & IP Protection
Status: Access via Request Only
Effective Date: December 2025
Related Project: National Art Fund (No: 2025-2-05-077-608)
Notice to Reviewers and Researchers
Thank you for your interest in the HITL-Lacquer LoRA model.
While our initial intention was to provide open access to the model weights (.safetensors), a post-submission data governance review has necessitated a restriction on direct public distribution. This decision adheres to the intellectual property protocols of the National Art Fund project and respects the agreements made with the contemporary artisan studios that contributed to our training dataset.
[cite_start]As detailed in the Ethical Considerations (Section 7.3) [cite: 391] [cite_start]and Data Availability Statement [cite: 486] of our manuscript, the training data contains proprietary high-resolution scans of non-public lacquerware artifacts. There are valid concerns that public diffusion model weights could be exploited to reconstruct or "memorize" these protected motifs (model inversion risks), potentially infringing on the artisans' IP rights.
How to Access for Reproducibility
To support scientific reproducibility while maintaining ethical compliance, we provide access to the model weights upon reasonable request for non-commercial academic research.
Please direct your access requests to the Corresponding Author:
- Prof. Yixin Li
- Email: 70201@hutb.edu.cn
- Subject Line: [Request Access] HITL-Lacquer Model Weights - [Your Institution]
We appreciate your understanding as we balance open science with cultural heritage protection.
Model Card for HITL-Lacquer LoRA v1.0
This model card provides documentation for the LoRA weights adapted for the Human-in-the-Loop AIGC system for lacquerware design, as described in our manuscript, "Enhancing Lacquerware Design Innovation."
Model Details
- [cite_start]Developed by: Jun Liang, Yixin Li, Zhileng Xiong, Qing Huang [cite: 2]
- [cite_start]Affiliations: Huangshan University, Hunan University of Technology and Business [cite: 3, 4]
- Model Type: Latent diffusion model fine-tuned with Low-Rank Adaptation (LoRA).
- Base Model: stabilityai/stable-diffusion-2-1
- License: CreativeML Open RAIL-M
- [cite_start]Funded by: National Art Fund project "Coromandel Lacquer Screen Creation Talent Training" (Project No: 2025-2-05-077-608) [cite: 472]
Intended Use
This model is intended as a creative partner for artists, designers, and researchers in the field of lacquerware. Its primary uses are:
- Design Ideation: Rapidly generating a diverse range of culturally and stylistically coherent lacquerware design concepts.
- Creative Exploration: Assisting users in overcoming creative blocks by suggesting novel combinations of traditional motifs and techniques.
- Educational Tool: Helping students of design and craft to learn and explore the visual language of traditional lacquerware styles.
Out-of-Scope Use
This model is not intended for:
- Autonomous Generation of Final Products: The model generates 2D concepts and is not designed to produce final, production-ready outputs without significant human intervention.
- Forgery or Deceptive Authenticity: The model should not be used to create forgeries of historical artifacts. All AI-assisted works should be appropriately attributed.
- Commercial Use Infringing on Cultural Heritage: The model should not be used in a way that commodifies or decontextualizes culturally sensitive motifs without respect for their origin communities.
Limitations and Biases
- Data Bias: The model's knowledge is derived from our curated 3,200-image dataset, with a primary focus on East Asian (Chinese, Japanese, Korean, Vietnamese) lacquerware traditions. It will perform poorly on styles not represented in this dataset.
- Performance on Rare Techniques: The model exhibits lower fidelity when generating designs that use rare or underrepresented techniques in the training data, such as tsuishu (carved red lacquer). This is a direct consequence of data imbalance.
- Complex 3D Forms: As a 2D model, it struggles to generate coherent surface designs for highly complex or sculptural 3D objects. The generated patterns may not wrap realistically around complex geometries.
- Stylistic Inconsistency: Attempts to blend stylistically or culturally disparate traditions may result in aesthetically inconsistent or incongruous designs.
Training Data
The model was fine-tuned on a private, proprietary dataset of 3,200 high-resolution images of historical and contemporary lacquerware.
- Sources: Museum archives, contemporary artisan studios, and exhibition catalogs from China, Japan, Korea, and Vietnam.
- Preprocessing: All images were resized to 512x512 pixels. [cite_start]A material-aware specularity reduction algorithm was applied to mitigate photographic highlights, using the formula $I_{corrected} = I_{raw} \cdot (1 - \alpha M_{spec})$ with $\alpha=0.7$. [cite: 184]
- Annotation: Each image was annotated with the multi-layer schema detailed in Part II of the supplementary materials.
Training Procedure
The base Stable Diffusion v2.1 model was fine-tuned using the LoRA technique. The training objective combined the standard diffusion loss with a perceptual loss term ($L_{perc}$) based on a pre-trained VGG-19 network to better capture lacquer-specific textures. [cite_start]The total loss was $L_{total} = L_{diff} + \lambda L_{perc}$ with $\lambda=0.3$. [cite: 130]
Training Hyperparameters
| Hyperparameter | Value |
|---|
| base_model | stabilityai/stable-diffusion-2-1 |
| model_type | Latent Diffusion with LoRA |
| lora_rank | 64 |
| lora_alpha | 64 |
| learning_rate | 1e-4 |
| batch_size | 16 |
| num_steps | 12,000 |
| loss_function | L_diff + 0.3 * L_perc (VGG-19) |
| optimizer | AdamW |
| precision | fp16 |
| hardware | 4x NVIDIA A100 80GB GPUs |
Evaluation Results
The model's performance was evaluated against baselines using quantitative metrics and a user study involving 12 participants (6 artisans, 6 students). The key results from our manuscript are summarized below:
- Fidelity (FID): Achieved an FID score of 12.7 on a held-out test set, compared to 18.3 for the vanilla Stable Diffusion base model.
- Diversity (LPIPS): Demonstrated a 22% increase in the LPIPS diversity metric compared to baseline generative methods.
- Controllability (Contour Adherence): When guided by user sketches via ControlNet, generated designs achieved 89% contour accuracy.
- Usability (SUS): Received a mean System Usability Scale (SUS) score of 82.4, indicating high user satisfaction.
- Qualitative Feedback: Participants reported that the system served as a powerful "inspiration partner," helping them amplify creativity and improve workflow efficiency without compromising their artistic agency.
