For now, the best would be to finetune an uncensored Qwen3-vl model. There are pre-made google collab to easily do this, but it is required to create a dataset of images with their expected spicy description.
