Hentai_Concept_test_z_image turbo

Incorporate concepts related to Hentai, utilize the AI-toolkit for training, and enhance the Z-image turbo model using 10,000 training steps, a fixed learning rate of 1e-4, and 1308 images mixed with 1024 and 1536 resolutions at various aspect ratios.

This is solely for testing purposes, mixing different prompts in both Chinese and English. The training loss heavily fluctuates back and forth between 1e-1 ~ 5e-1. The base model might be relatively deficient, but its fundamental understanding is not poor.

I hope you will like it, even though it might not be very diverse.

___________________

Since the training was conducted without LLM augmentation to generate more detailed prompts, the prompts used were relatively short, leading to strong randomness. Moreover, over 50% of the concepts basically failed to train, which is likely due to a lack of knowledge or issues related to the model distillation. For months, or even longer, it has been observed that prompt length significantly impacts training. Structurally ordered, highly consistent, and precise prompts are a critical key to successful training. However, this requires substantial manual effort, even with a VLM (Visual Language Model). Increasing the Rank from 32 to a higher value is possible, but this would lead to excessive memorization/over-reconstruction, which is not the desired effect.