I think the problem is the ratio. My PC run out of vram before it can finish so I usually stay at 512x512. The preview are in 1280x1920 maybe that's the problem.

maybe your img is too small,try bigger size,because mine are really good,but please don't forget the embedding he've use,it's in here https://civitai.com/models/4514/pure-eros-face

@bishop Can I see one of the image you made? Like I said, I generate it in 512x512 but upscale it to 1280x1920 as the original. It doesn't look as good as the example.

In this render, from the users comments, the face appears much better:

https://civitai.com/gallery/114624?reviewId=15456&infinite=false&returnUrl=%2Fmodels%2F9871

Is it just a matter of prompting or something else ?

@ritcher1 Unsure, I'm just using the exact metadata as the preview and other people around. I really think you need to generate it at higher resolution.

you need the embedding in his comment

Add this if you want the same face, pure-ero-face: https://civitai.com/models/4514/pure-eros-face

@bishop Sorry fort the late reply and the spam. The site has been slow for me and I didn't see my comment pop up. Yes, I do have Pure-eros-face and bad artist embedding.

me too,and I do have that two embedding files

same here, don't know what im missing

check this video related to this Lora https://www.youtube.com/watch?v=ZVe310AEz_Q

cannot replicate the results either.... I even fired up a colab for it, same results, I drag the .png to make sure the parameters are the same...still.....

@cdan Model Creator is probably using a different GPU or using xformers. Example I have an RTX 2060 and you have RTX 4090 the results would be different because we have the different amount of Cuda, Tensor, Vram, etc.

@adrien1207 thank you for your reply, I figured out the problem, and now I can reproduce 100%.
The images are generated using automatic1111, but with the option "Do not make DPM++ SDE deterministic across different batch sizes.". Also people are generating images using negative prompts with embeddings, but do not have them present in the embeddings folder, this happens even with Lora's. You have to figure our what exact combination they used. Also there are some other extensions e.g. sd-webui-tome which makes generation faster but gives totally different images with same prompt.

A way to fix all of this is for automatic1111 to put, into the .png metadata, all of the settings including version of the UI itself, also errors and used embeddings etc...