@Meina sure

did you download the lora, embeddings and put them in the correct folder?

@freedomgiver11 i am not sure, i just run the webui to use cpu resource to render images by adding the extra command line args: --use-cpu all --no-half

Have you checked clip skip, ENSD, resolution and variation seed?

@roninjiang You need to download the same checkpoint, lora, embeddings, in order to get a similar image. Those items teach the ai how to draw.

@freedomgiver11 i think it uses the same checkpoint, lora, embeddings when choose the CPU target. In my opinion the difference is the actual calculation is moved from GPU to CPU.

Of course, running on CPU is extremely slow.

@Meina I found that the body types generated by meinamix are very consistent. It always generates mature women when it sees words like "breasts", with big buttocks and long legs. Is there a way to favor girls?

By the way, the cool colors I refer to are more inclined to color styles, v7 seems to prefer cool colors such as blue and black.

@Meina For convenience, I put the example here. You can clearly see the effect of "breasts" on the legs, it tends to output wider thighs regardless of breasts size.; the model tends to have long legs, sometimes it is exaggerated, and I have no way to control it, which affects the generation of girl characters; I lack some methods to control the body shape. https://we.tl/t-VhxtxU3UxB

@Meina Here is an example of using v8 to reproduce the v5.1 sample image, you can easily find the corresponding image according to the image parameters. From my point of view, v7+ clearly prefers to use cool colors (black background, blue sky) in the part where the prompt indicator is missing. https://we.tl/t-D6yu65TMNy

Using 40 custom models to overcome the limitations of SD2.0 doesn’t sound like a fun time. SD1.5 has the ability to create the most unique and gorgeous styles by combining the styles of different artists. My interest lies not in celebrities or pornography, but in the potential of an AI model with comprehensive knowledge. Such a model could advance image intelligence by improving its understanding of anatomy through nsfw material and generating unique styles by synthesizing tokens from different artists. The prompt, whether it be "Greg" or "Disney," contributes to a general understanding of the visual world, rather than serving as a means of copying. Removing certain elements from the model hinders its ability to create new, original art styles and diminishes its usefulness for various purposes. This includes both uncreative and creative uses, making them less powerful overall. Until 2.x can provide something at least as good as that, I'm not interested

Interestingly, the model insists Miku has tattoo on the right arm, or Ganyu has two horns on each side. Removing the character's name in the prompt then repainting did the trick for me.

Add a review

WTF? That site is a thief of model makers' hard work!

你好，能说说是怎么复现的吗。我连第一张sample(黑衣短发女生)都没能复现出来。我在SD-WebUI使用这个ckptmerge模型，把第一张图的各项配置都复制了一遍，然后添加了作者在评论区说的kl-f8-anime2这个VAE，分辨率我选择了576*1152，其他的settings我都没有修改。最终生成结果感觉相差很多，人物发型和胸前的蓝色圆圈都不一样。

Hello, could you please explain how to reproduce the results of this model? I couldn't even reproduce the first sample (short-haired girl in black). I used this ckptmerge model in SD-WebUI, copied all the generation data of the first image, and added the VAE kl-f8-anime2 that the author mentioned in the comments. I chose a resolution of 576*1152 and didn't modify anything in "Settings". In the end, the generated result looks very different, with a different hairstyle for the character and a blue circle on her chest.

更新

很抱歉打扰了，我的问题已经解决了。我重新将分辨率设置为512*1024，结果就和样例几乎一致了。另外，我看到作者说VAE不用另外下载，而是已经附带在这个模型当中了。

update

sorry to bother you. My problem has been solved. I reset the resolution to 512*1024 and the result is almost the same as the sample. Also, I saw the author mentioned that VAE is already included in this model and there's no need to download it separately.

@AIzab glad to see that .have fun!

@Meina Thank you very much, and thank you for your work

I had a bit of success with painting over them in photoshop, then importing into img2img and slowly rendering it from low denoising strength, then gradually increasing it. I was able to get more skin show up instead of the collars, but the model definitely wants to put collars there.

also: I don't get it, what means old meina version - without vae but at v8 it's don't work with loras normally with none and with some good vaes

Aaaaaand... just noticed, that almost every other anime-style checkpoint has the same thing in common about hairs.

I've used "forehead" in positive and "hair between eyes, bangs" in negative, with ((emphasis)) as necessary, when I noticed this on ponytails, seems to work at least a good percent of the time

@dizzysix "sidecut" and "punk" also seem to work, but on higher weight it makes the hair alwas short and blue/purple... and sometimes messes up the whole face :D

@indgo FYI that's just a trope in anime. Asian peoples tend to have straighter, courser hair, on average, than westerners, and it is (or was) especially popular in Japan for young women to comb the top of their hair like this. I want to say it started to become prevalent in anime and manga after the 90s? You don't see it in ramna 1/2 or sailor moon. Other anime conventions like the giant sideburns have been around for much longer. I wanna say it was popularized in visual novels before spreading to manga and anime but don't quote me on that. Fun fact: The western developer of Doki Doki Literature Club worked hard in his self-described "programmer art" to emphasize this combover style explicitly. It's possible it existed in IRL fashion and was already in manga and anime in the 90s or 80s or even before-- just in such an understated way, casual western viewers wouldn't notice it-- and that only recently it may have become in vogue to emphasize it.

I'm guessing, though. I'm just a casual anime fan, not an anthropology major.

@TranscendentThots tbh, I'm watching a lot of anime. I don't have a particular problem with the buns, or this style, but what I see here is a very-very specific style. If you compare 4 (non-lora) generated character hairs and 4 random anime character hairstyles, you will find (at least I do), that the generated ones show a very hard similarity with eachother, while vanilla anime characters (be new or old) show much more diversity. And with this, its pretty easy to tell if such character is generated, because the hairstyle is so distinct, that I did not managed to find an anime character with this exact hairstyle (I tried to find the source of this style, or which component of the model mix introduced it). But it might bs just me... some of my friends see no similarity between the characters.

Most hair will look the same if you don't apply a hairstyle. use a hair wildcard and you will see allot more results

@sownheard Wait, what's a wildcard? You can just type something like random hairstyle, random skirt length, random gender, or whatever, and it will give you more variety in the outputs than leaving it to the default? That sounds incredibly powerful and useful. How do you do that?

This is a known issue with some models, possibly related to automatic1111 webui. Switch to another model, generate a couple images, and switch back generally resolves -> https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/6923#issuecomment-1397768771

@tensaishun Yeah I saw that post, but my situation is that even switching another model doesn't work when this model is present in model folder. Only when I moved this model to trash will solve it...