By downloading this model you denounce Zoe and endorse ANNIE as best league girl.
Previous Annie loras are a bit old and stiff, so I updated my training data from when I made my Annie embed and trained a lora with settings that I know to work pretty well. It seems to be an improvement over previous Annie loras for the things I tested it with. 113 training images, no tag pruning.
If you want stuffed animal tibbers, add: stuffed animal, teddy bear, stuffed toy, object hug. The model isn't really trained for big demon mode tibbers but I guess you could try adding: animal, bear, and other things.
If you want fire, add: fire
Otherwise put those in your negative to avoid them creeping in.
Description
FAQ
Comments (7)
I always run a test generation of any Lora, and I noticed I am unable to achieve a similar picture, let alone style, of your example pics. I got my hands on the same model, copied over the prompts and the seed, set the steps, and clip steps accordingly, yet to my surprise, the art it makes is nowhere near it. Here is the pic I generated, and I was trying to generate this one.
,%201girl,%20solo,%20a%20beautiful%20masterpiece%20portrait%20of%20a%20girl%20,%20standing,%20contra.jpeg){kind=link}
png info isn't showing a lora hash for your image so you either aren't actually using it or you're not using up-to-date webui when generating your metadata. Looking at the generation data of your provided pic, I notice that you invoke the lora as <annie18>, which is incorrect unless you renamed the lora to that. Interestingly, I made an Annie TI embed some time back with the name Annie18 (since its vector size is 18), which makes me think you might have mixed things up. So I'd double check that the lora is being referenced correctly and you've downloaded this lora (which I believe ought to be named "Annie1024Test".
Beyond that, check that you have a vae on (your image is washed out and your generation data doesn't show one), your clip skip should be set to 2, and that you actually have the negative embeddings at the start of the negative prompt downloaded and in your embeddings folder. If not, simply remove them.
@wheelinghubcap I have made a tremendous amount of errors, starting from not realizing to send you all the info you need, to not knowing what a vae is and, somehow, not having the right Lora on. I won't deflect with 'oh I just started using SD', sincere apologies for wasting your time, and sincere thanks for this illuminating response. I got my hands on the embeddings, a vae, and everything else, and now I am getting very similar results as yours. This was thoroughly informative and helps me with future endeavors, again, Thank You.
@rth90283552 You are not wasting my time at all and I am happy to help! I'm glad you were able to get it working, let me know if you have any other questions.
@wheelinghubcap I actually do have one, how did you find out which lora I invoked? How did you find the generation data, how is it embedded and how can I check it?
@rth90283552 in webui, use the png info tab. The image has to be downloaded from a site that doesn't automatically remove metadata. A lot of websites automatically remove metadata because it can have things like location info from cell phone pictures.
نشط: يدخل Jax في Evasion ، وهو موقف دفاعي ، لمدة ثانيتين ، مما يتسبب في تفادي جميع الهجمات الأساسية غير البرجية ضده طوال المدة. يحصل Jax أيضًا على تقليل الضرر بنسبة 25? ، مما يقلل الضرر من جميع قدرات منطقة التأثير التي يتم الحصول عليها من أبطال البطل. يمكن إعادة صياغة Counter Strike بعد ثانية واحدة ، ويتم ذلك تلقائيًا بعد انتهاء المدة.
RECAST: يلحق Jax ضررًا جسديًا لجميع الأعداء القريبين ، ويزيد بنسبة 20% لكل هجوم يتم تفاديه ، وزيادة تصل إلى 100? ، ويصعقهم الصاعقة لمدة ثانية واحدة.