Note:

This is a LYCORIS/LoHa model, and you need to install a1111-sd-webui-locon extension to use this model.

And don't use sd-webui-additional-networks to load this model.

Place the model in the path '.\stable-diffusion-webui\models\Lora' to use it.

Trained using 500 pointillism paintings, sourced from artists such as Georges Seurat.

Based on novel-anime-full model.

Trigger word:pointillism

Suggest weight:1-2(Adjust the weights appropriately when you think it is necessary)

It should be noted that this model has several issues. Firstly, it cannot guarantee good performance on many models. I have tested it and found it to work well on the Anything and NAI models.(In the previous tests, I experimented with using stable diffusion 1.5 for training, but the results were not very good.)

Secondly, this model mainly relies on the 'pointillism' tag to operate, and its effectiveness depends on the degree of support the model has for this tag. Additionally, for excessively complex prompts, the influence of the style on the generated images may be weakened. (I speculate that this is because during training, the model did not directly fine-tune some of the weights of the base models, resulting in underfitting in some areas, which in turn affected the style of the generated images).

For example, if I write '1girl,white hair,blue eyes' in the prompt, the LORA style does not fit these tags, which results in the generated character resembling the style of the base model. Therefore, I conducted a test by replacing the CLIP captioning with the wd14-tagger. The result was that the model learned the style of the dataset when generating characters, which was similar to post-impressionist paintings, but it did not seem to learn 'pointillism' compared to the CLIP caption version.(But you can still use it as an oil painting style model.)

Thirdly, it does not support merge with other LORAs to use, as even with a character LORA, there is still some influence from the trained style, which can cause the style of the merge LORA to be compromised.

If you have any better training suggestions to address these issues, please leave a message for me. I would greatly appreciate it. Similarly, I will upload the dataset. If you are interested, you can download or improve it for training.