I'll be brief.

20 steps of LCM + KL_OPTIMAL at 1MP+ is a resonance point where the LCM consistency trajectory aligns with the KL_OPTIMAL frequency grid and the distilled structure of Krea 2 Turbo INT8.

This method only reveals its full potential on well-trained models. If the LoRA was trained in the standard way, you won't see a meaningful difference from the usual 8-step setup.

Based on my observations, this point is reached at around 6,000 training steps with a learning rate of 0.0001. Once this threshold is crossed, the model begins to learn lighting, shadows, and minute details, effectively overriding the base model's behavior for those elements.

I expect new models to appear on this page quite infrequently, as a few unsuitable images in the dataset will negatively affect the result—which only becomes visible after 6,000 steps. After that, I rebuild the dataset and start over.

Once the LoRA has been successfully trained, I edit all the blocks to reinforce those that have effectively internalized the information and to weaken those that have absorbed unwanted information.