My First LoRA for QWEN-Image: Anime-Style NSFW Illustration

This is my first published LoRA and my first attempt at creating one for QWEN. It's an experimental model focusing on generating anime-style NSFW illustrations.

Since dedicated illustration-style Loras or fine-tuned models for Qwen-image are currently scarce, I decided to experiment with creating an NSFW LoRA. This initial version is trained to recognize basic sexual positions and concepts from Danbooru-style tags.

Used Prefix:

anime-style illustration

Current State

This is a proof-of-concept release with significant limitations:

• Small Dataset: Trained on only ~200 images.

• Image Quality: Not all images has good quality and has artifacts.

• Captioning: Captions are diverse but not yet optimally structured.

Known Limitations & Issues

Due to the above constraints, expect the following:

• Poorly Rendered Sex Scenes: Intimate scenes often lack coherence and messy.

• Inconsistent Clothing: Character outfits may not reliably follow prompts or concepts.

Despite these flaws, it can still be used cautiously for:

• Generating character concepts in an anime and lewd style.

• Applying a base Illustration in generation.

Training Details

• Steps: 12,000

• Base Model: Qwen-image 2512

• Use Prefix in prompts: anime-style illustration

Future Plans

The next steps to improve this model are:

1. Curate the Dataset: Remove low-quality or irrelevant images.

2. Expand the Dataset: Increase the training set to at least 500 high or decent quality, well-captioned images.

3. Refine Captioning: Implement more structured and accurate tagging, currently, the sentences are generally complex and multi-clausal.

This cleanup and expansion process will likely take 1-2 weeks.