I actually didn't plan on releasing this model originally, not until I kept getting many questions and requests. Curiosity got the best, I guess, right? People always want what they can't have.

I never released it because it was an experimental Lora I used to test Qwen's adaptability to large datasets (like, >=1K images) right when Qwen-image first came out. Its whole purpose was to give NSFW capabilities to both the Qwen-mage and image-edit models. It worked, but honestly, it didn't quite meet my standards.

My idea was to gather some intel for future NSFW checkpoint fine-tuning – the checkpoint is my real goal. Ultimately, I'm hoping to be able to do this with a training set of over a million images. But with money and time being an issue, it honestly sounds like an unrealistic pipe dream right now.