Z-Image Turbo Info:
Sampler Settings:
Use Beta Scheduler for best results, I've also been using Euler for the Sampler with good results.
PROMPTING:
Based on my tests with Z-Image, please use the word "vagina" instead of "pussy" to show a vagina.
There will be some oddities with some anatomy as I continue learning about training this model, so any feedback would be great!
Flux Dev Info:
This model excels at nude female anatomy. Its not perfect in this first version, so several generations might be needed for certain things. Some things to note, this REALLY wants to make nude women, so if you're trying to add clothes, I suggest lowering the strength a bit to lower the models tendency to make nude women. Second, I think there weren't enough asian women in the training data, so it struggles to make asian women. It will do it, but in the second version I will add more asians.
SD 1.5 Info:
This is a model to give your subjects an oiled/wet skin look. This is meant for photo real images, but I'm sure it will work on anime as well. No keyword is needed, you can just keep the weight at 1, or lower it if needed. It's very easy to use. The only thing I will suggest is to use Watermark as part of your negative prompt. The images I used had some watermarks, but I will release a v2 soon with these watermarks removed. This model is also fairly nsfw, so if you don't want nude images, you will likely have to negative prompt for your liking.
Share your images! I'd love to see what you make!
Description
FAQ
Comments (10)
Anime works pretty well too, I used a weight around 0.7.
I find that in general it doesn't really want a weight higher than 0.6 or maybe 0.7, and this needs to be reduced a bit if other LoRAs are involved. But if you keep the weight low to middling it works really well.
any prompt to activate or?
no phrase needed! You can even check the prompts I used on these images to see
@MyNameIsThis ok, i will upload soon some test, i love the way the skin looks
with loras, don't think of trigger words as on/off switches. think of them as descriptors that tell the AI what you want it to do with the lora. if you don't use the trigger word, the AI will simply generate images similar to what it was given during training according to its understanding of the lora subject. by adding a trigger word, you can tell it to use the lora in more specific ways.
for example, say the lora is a hat with the tigger word "lorahat" and you want an image of someone holding the hat in their hands instead of wearing it on their head. if you add the lora to the prompts without the trigger word, it may only generate images with the hat on peoples heads. but by using the trigger word in a prompt like "character holding lorahat in his hands", the AI will have a better idea what to generate
@rjox When I first started training Loras, I was using the Blip method of captioning, and I'd start every single description with the trigger word. However, I recently moved over to WD14 captioning, and the process has been much easier and yielding me better results. It's also been easier to trigger the Loras. I've been creating private Loras of people, and I just have to use the word "woman" or "man" in the description of the scene, and it will use the model perfectly.
@MyNameIsThis Yeah for sure. Wd14 is my favorite too for non-characters
not work.
can you give me more information? Clearly there are hundreds of examples of this working