FP8 + Lightning 8-steps Lora is recommended. [ CFG 0.9, 1.0 , 1.3 max ]
By downloading the LoRa, you agree that you are fully responsible for how you use it with Qwen-Image-Edit.
Eva V3
Just a new look for it.
Qwen-Image-Edit workflow in the zip file. ( SDXL image sizes are not compatible with Qwen-Image-Edit)
NSFW is limited to a position.
trigger word for Qwen-Image-Edit: in a kneeling position
Example prompt: (don't use oral sex)
a woman is in a kneeling position and performing oral penetration. she is performing fellatio. The man's erect, penis is in her mouth.Keep in mind that it is trained to serve other purposes. The NSFW content is limited to that act, and it is not by any means 100% accurate. The appearance of male genitalia is greatly affected by Qwen-Image itself, and it is not something that a LoRA of this caliber can fix.
For me, the idea was to see how Qwen-Image-Edit performs with NSFW prompts.
Eva Gothic style
Trigger words:
For anime: anime style, Eva_gothic.
Semi-realism: Eva_gothic.
Neutral: no trigger word, LoRA strength: 1.0
The dataset is generated by the AnimaPencilXL checkpoint and Gothic Girl to the point where most of the images had some sort of deformation, so I decided to use another source. The dataset is a combination of a black-blue theme, neutral tones, and semi-realism. It is mainly another step for me to get to know Qwen-Image better.
The captioning of the dataset was done in two forms:
anime style, Eva_gothic for images with a gothic look and feel.
Eva_gothic for semi-realistic images.
Without trigger words and LoRA stength 1.0: for images with different colors and themes, such as modern day urban settings and cyberpunk, ect.,.
Depending on how it performs, I will try my best to improve it in the next version, if it proves to be necessary.
https://civitai.com/models/261336/animapencil-xl
https://civitai.com/models/248965/gothic-girl
Eva Qwen
This character is related to the article below:
Basic Guide to Qwen-Image LoRA Training
https://civitai.com/articles/18998
However, it can still be used for generating images with consistent characters.