Multi-Style LoRA for Qwen Image 2512

Hello everyone! I’ve been experimenting with a Multi-LoRA concept for the Qwen Image 2512 model. The initial results are quite promising for a proof-of-concept.

Style Guide & Trigger Words

This LoRA features 4 distinct styles. You can experiment with them to achieve different aesthetics:

• High quality 3D render: Provides the most polished results.

  • Note: The dataset contained some watermarked images. If you don't prompt for specific text or icons, the model might occasionally generate its own "watermarks."

• Anime 3D render: Functional, though still being refined. The output differs significantly from the "High quality" version, so try swapping between them to find your preferred look.

• High-quality anime illustration: This style was trained on a custom dataset of high-quality generated images.

• Medium-quality anime art: Interestingly, this sometimes results in better anatomy and cleaner lines than the "High-quality" tag above.

Recommended Settings

To achieve the results (as seen in my examples), use the following configuration:

• Lighting LoRA: Use a 4 steps lighting lora

• Sampler: euler_ancestral

• Scheduler: beta57

• Resolution: Defaults to 1024×1024

The training process:

• Dataset Size: 900 quality and some good images. Yeah was many futa.

• Total Steps: 22,000

• Resolution: 1024×1024

• Precision: fp8

• LoRA Rank: 64

• Compatibility: Also partially works with Qwen Image Edit 2511.

Roadmap & Future Plans

I believe that a cleaner, larger dataset will yield even more incredible results.

• Upcoming: More refined 3D and Anime concepts in the NSFW category.

Feedback

I would love to hear your thoughts!

• What should I add next?

With your suggestions—I'm open to many kinds of feedback and kinky things!