Second Training Run of the Prepared Dataset for Qwen-Image 2512

I’ve completed a second training run using the prepared dataset for Qwen-Image 2512.

The dataset included 400 images that I created and enhanced in Illustrious.
This time I tried using rank64, which is why the LoRA file is so large. I also used a high number of steps — 20,000 — which I think was justified for such a small dataset, judging by the tests I ran between checkpoints. I’m more than satisfied with the current results.

⚠️ Important: Recommended Settings

If you decide to test this version, I used the Qwen-Image-2512-4steps-V1 base. For the best results, use the following configuration:

• KSampler Steps: 4-6

• Sampler: euler_ancestral

• Scheduler: beta57

These settings appear to work best for maintaining possible stability and quality with this specific LoRA.

Current State of the LoRA

The LoRA is still in an active development phase, so there may be artifacts and failures to understand many concepts.

Right now it’s a mix of multiple concepts, but roughly 60% of the dataset is either fully nude characters or characters in very revealing outfits.

There are very few sexual acts and anus-focused images, so the model often makes mistakes in those areas.

Concept Gaps and Dataset Quality

There are still significant gaps in concepts such as:

• Sex positions

• Different types/styles of clothing

The current dataset and captions are not perfect, but they already provide decent training results.

Future Plans

There’s still a huge amount of work to be done, especially if we focus on the main concepts. I plan to keep working on this, gradually improving both the dataset and the LoRA.