Introduction
Our Chinese pattern collection LoRA model dataset comes from Shuge - Chinese Pattern Collection, containing nearly 100 traditional Chinese pattern flat images. After detailed image processing, we obtained a LoRA model with considerable background understanding and generalization capabilities.
Model Download
- Selected models: https://civitai.com/models/29858/chinese-traditional-pattern
- Complete models (30+): latentcat-models - Hugging Face
Usage Guide
In the training set annotations, there are five main phrase categories:
traditional chinese pattern
Classification word. Adding this word makes it easier for the pattern to become a background for other images.
with circular symmetrical composition
with square symmetrical composition
with special-shaped symmetrical composition
Composition classification words. Choose between circular/square/special-shaped to more easily generate circular/square/irregular compositions.
red and blue color scheme
Color preference. You can select 1-4 colors, with earlier colors having greater weight in the image. This can be understood as these colors becoming the first visual impression of the image's color distribution.
featuring horses and clouds
featuring flowers and branches
Feature descriptions, such as featuring flowers or featuring lotus and lotus leaves. Flowers and clouds work best, while other terms have slightly weaker generalization ability.
a flower in the center
Describes central features, like a flower or lotus and lotus leaves as focal points. Flowers and clouds are the most effective terms, while other terms show somewhat weaker generalization.
You can expand upon these five fixed phrase patterns. For background use, you can lower the weight and add "1 girl".
Layer Analysis:
As seen in the layered images, under high weights, by only enabling LoRA layer weights like MIDD and OUTD, you can better preserve pattern styles in character background images compared to simply reducing weights.
Recommended Model Versions:
Here are some model versions with distinctive effects. However, image generation results depend on parameters, seeds, and other factors. Feel free to experiment, and we welcome your comments.
Featured Model Versions
Here are some featured model versions:
- 4: Characterized by more cartoonish patterns and softer colors
- 7: Better keyword fidelity, more aligned with dataset tendencies
- 10: More symmetrical, tends toward non-center-symmetric compositions
- 13: Softer colors compared to 10, shows a third scattered composition at high weights
- 15: Features high saturation and horizontal symmetric composition
- 16: Characterized by high saturation + scattered composition
Character-Compatible Versions
The following versions work well with character generation:
- loha01_old3: Better at cartoon lighting, strong decorative composition and patterns
- loha04_old3: Easily generates full-body compositions, good lighting preservation
- loha06_old1: Less compositional constraints, natural character poses
- loha02_old2: More background layers, strong symmetry
Note: If character generation is extremely difficult, use the Lora-Weight-Block plugin for layer adjustment, enabling only IND or OUTD layers for easier character generation.
Dataset annotation and LoHa training methods are still being updated.
Image Processing
Image Tagging
Tagging Process
Using BooruDatasetTagManager to load folders, classifying the cleaned dataset with qualifier words for background, color, style, and composition, using related similar words for training
Test Outputs
Issues and Limitations
The dataset has undergone extensive manual screening and has lower generalization ability for asymmetric patterns. It needs dataset supplementation or continued training. If you have more suggestions, please comment or submit a PR.