Examples
Before we go any further, it makes sense to demonstrate what results can generally be achieved.
Danila Poperechnij | More
Karina Istomina | More
Lana Del Rey | More
Skryptonite | More
It's worth noting separately that these are the results generated by the previous generation of text-to-image neural networks based on Stable Diffusion 1.5 (SD1.5). The current generation of neural networks based on Stable Diffusion XL (SDXL) demonstrates a completely different level of quality. However, these neural networks are much more hardware-demanding, and the algorithms of their training and image generation are far from optimal at the moment. Therefore, right now I can recommend delving into them only to enthusiasts. Nevertheless, it won't hurt to demonstrate what they are capable of.
Karina Istomina | More
These are the results obtained using the base SDXL checkpoint and the LoRA model trained on it. Meanwhile the base SD1.5 checkpoint produces something completely inappropriate.
It's frightening to imagine what will come next, considering that less than a year has passed between SD1.5 and SDXL. However, training models on SDXL is fundamentally no different from training models on SD1.5, so everything we will discuss further is applicable to both.
Next - Dataset Preparation
