I2I always generates nipples that are not present in the original image.
I use anime images to convert them into real-person images, but I've noticed that the generated images always show nipples on the breasts, even when using negative prompts. Furthermore, if I use an SFW model, it still generates spots resembling nipples,like this https://img.bian666.cf/file/1763043471812_Snipaste_2025-11-13_22-17-27.png
Is there any way to avoid it?
"SFW" models are trained on breasts and buttocks
I use anime images to convert them into real-person images, but I've noticed that the generated images always show nipples on the breasts, even when using negative prompts. Furthermore, if I use an SFW model, it still generates spots resembling nipples,like this https://img.bian666.cf/file/1763043471812_Snipaste_2025-11-13_22-17-27.png
Is there any way to avoid it?
Could you share your workflow? I'm curious how to do the anime->real and didn't know the AIO could do it.
I use anime images to convert them into real-person images, but I've noticed that the generated images always show nipples on the breasts, even when using negative prompts. Furthermore, if I use an SFW model, it still generates spots resembling nipples,like this https://img.bian666.cf/file/1763043471812_Snipaste_2025-11-13_22-17-27.png
Is there any way to avoid it?
Could you share your workflow? I'm curious how to do the anime->real and didn't know the AIO could do it.
You just need to add a LoRa module: https://civitai.com/models/1934100/anime2realism.
Yeah i'm aware of that lora but i'm not clear how to just have a workflow for anime to real conversion, and i went through that civitai page and the posts and wasn't able to find a workflow link, just a handful of people also asking for the same.
i assume you add a loraloadermodel only node between the model load and shift (or sampler). In whatever workflow you use. and there you select any (Qwen)lora you want.
the same as in any non aio (and non Qwen) workflow.
it goes directly after your model loader. you can chain them. multiple lora loaders in a row and then you use the last lora loader output (model) and continue as it was before.
Not sure if a 4step load should come before or after a people/style lora (heard different opinions so it probs doesn't matter) , but this is irrelevant here since that is integrated in the aio model already.