V3.3 Glittering gravel：
Key Technical Changes
Dataset Expanded: The dataset has been increased to over 9,500 images.
New Training Method (LoKR): I've switched from LoCON to LoKR for this version. The most immediate benefit is a much smaller file size.
Noise Offset: This version introduces noise offset, which should provide a noticeable improvement in image contrast.
Dataset Refinements:
To help reduce the smooth, plastic-like "AI feel," I've added a selection of real-world photography (without portraits) and removed images generated with the Illustrious model.
Performance and Usage
Quality and File Size: Don't worry about the smaller file size. In most cases, its performance in rendering lighting, details, and backgrounds is on par with previous versions.
Not a Style-Specific LoRA: Due to the training method and data variety, this LoRA does not strongly fit one particular art style or replicate specific characters perfectly. The advantage here is that it won't hijack your intended art style, giving you more creative freedom as long as you avoid overfitting.
Artist Strings: I have not observed any degradation when using artist strings. In fact, the overall results are often more flexible and diverse than in previous versions. This is likely because the LoKR training method reduces the model's rigidity, making it more versatile.
This version still excels at enhancing details and lighting, and now offers greater flexibility in style combinations.
Thank you for your continued support, and I hope you enjoy experimenting with this new version.