@ppkkmoon can pony and animagine 3 be mixed?

@ppkkmoon 说得跟你不是一样，哼！

@ppkkmoon everything is up to date, do I have to put the SDXL 1.0 and SDXL vae in the model folder? because I don't have those

@ppkkmoon Thank you so much, I love your Model, I have placed those checkpoints in my corresponding folders, I will update you, thank you for wonderful work <3

@ppkkmoon finally I got it to work thank you so much, I love your model so much <3

@ppkkmoon Thank you for reply!

I'm sure if you apply it with fooocus，the generated image will show washed out colors, but I think this problem may be related to fooocus itself.

@BIG_A in a sense, yes, I haven't found in fooocus how to choose VAE. But most checkpoints already have it baked in

https://www.reddit.com/r/fooocus/comments/1aivd1e/how_do_i_load_or_change_vae_with_fooocus/
OldFisherman8 said "Some of the SDXL models don't have baked-in VAE and will look washed out. One trick that seems to work is to place a similar type of SD 1.5 model as a refiner, set the 'Refiner Switch at' setting to 1, and change the 'Refiner Swap Method' in the Advanced/ developer debug mode section to VAE."

I used animePastelDream_softBakedVae.safetensors on Fooocus, it worked.

@nagaete289 it works, thank you! Nice trick

这不是xl版本吗？

@ppkkmoon Interesting, so its just additional keywords that the dataset had like the classic "masterwork" or "best quality"? Thanks for the reply.

我试过了，还是一样的，不知道是不是加入这些提示词（（8k，RAW照片，最佳质量，杰作：1.2），极其详细的CG 8K壁纸，（水晶纹理皮肤：1.2），极其精致美丽），还是因为没有加入大模型关键字

@2205412484237 你是不是用fooocus出的图？

加VAE可以的，很好的模型，是我操作失误了，

@ppkkmoon 瞧你酷的，唉！不升级就不升级吧

@ppkkmoon 不一定需要升级，显卡比较烂，求出个Turbo版本

@ppkkmoon I'm using a website to run your model so, maybe it's the website messing it up? however, been working fine until today. Maybe it's issue... vae anyways still love your model.

if it's grey but have a little bit colored, then you are drawing without VAE.

i have the same issue, using a VAE doesnt solve it. that just breaks the generator and its wayyy too over saturated

Ponydiffusion models seem to do sketch style or monochrome ALOT atleast for me. And I mean ALL Ponydiffusion models from mixes to sdxl. So maybe keep spamming the negatives with (sketch,monochrome,grayscale,greyscale).

I guess it's the model its self?

Seems like it's the model itself, I haven't had this issue with other Pony models.

@guy907223982 which other pony model?

@Cookcookies Both of these models work fine for me: https://civitai.com/models/288584, https://civitai.com/models/308079

I did change the vae and it started working so i'm not really sure why it specifically happens for this model.

Same. Worked the first time after switching from PonyDiffusionV6 to this and then images were gray. Setting VAE to automatic, none, or sdxl doesn't work. Switch back to PDv6 and it works. Strange.

Vae seems needed for sure, try the "low memory" one here on Civitai

If you click the link the provided in the description, you should see the included artists native to it.

@ppkkmoon Thank you!

Okay... where can I find it? xD

@Werescythe madebyollin/sdxl-vae-fp16-fix · Hugging Face

How do you make sure it uses the VAE?

@Werescythe It depends on the software/UI you're using. In Automatic1111 VAE menu is located in the options. Download an official VAE or a fixed fp16 VAE for SDXL, put it in the VAE folder, and then you can select it in the dropdown menu.

@Rousherte I'm using ComfyUI.

@Werescythe All the steps remain the same, you only have to add a node that loads a VAE.

With 6 GBs of VRAM you should be able to run this model as any other SDXL model. Try using --medvram and --xformers commandlines (SD Next, A1111), or trying other apps like ComfyUI, InvokeAI or A1111 Forge. You have to make sure that System Memory Fallback is disabled in NVIDIA Control Panel, systemwide or specifically for a Python executable. SDXL regularly hits 6 GB VRAM usage which by default prompts the system to fallback to RAM, reducing the speed tenfold.

@Rousherte I disabled System Memory Fallback in NVIDIA Control Panel and now works. Big Thanks!

For those also having an issue with grey/fuzzy screens. Search "SDXL VAE" on civitai.

This one worked for me but I doubt it's the only one that will work: https://civitai.com/models/296576/sdxl-vae

@ppkkmoon Okay, now let's say that I want to have two original characters and I'm using prompts for each one, how would I distinguish who has what prompts?

@Werescythe using the format 1boy (x, y, z), 1girl (a, b, c) helps but it still isn't perfect, like 30% success rate. I do that, pick the ones that are closest, then use inpainting to fix it from there. Sounds like SD3.0 they just announced today will be much better at this.

@malias693 Thank you, this was very helpful.

This might be a silly question, but us it possible to do it with two girls as well?

@Werescythe Yes, any combo of subjects. Still a low success rate but better than nothing.

@malias693 so would that be like 2 girls, 1 girl (x,y,z), 1 girl (a,b,c)?

@Werescythe I've been messing with it and am actually finding more luck like this... 2girls, 1girl, a, b, c, 1girl, (x, y , z)

@malias693 I'm just wondering, do there need to be spaces after the commas?

@Werescythe I use spaces but I'm not sure if it matters. Again, it's not a guarantee, you could run into good or bad luck. Inpainting with only the properties of the character you want in the prompt (so it could still say 2girls or whatever but only mention one hair + eye color etc) is the only sure-fire way I've found.

havent tried it on this model yet but this prompt setup example generally gives me good results

hdr, rich color, highly detailed, 2girls, yuri, adult, 25 years old, <starfire (teen titans), orange skin, submissive,bondage, tied up in ropes, collar>, <raven (teen titans), grey skin, short hair, unamused, bored>, strapon sex, strapon dildo, sex, fucked, bdsm, night, wallpaper background, path traced lighting, by zone

@ppkkmoon that will be amazing if it happens!

@ppkkmoon What's the SD 3.0 compatibility with previous models?
Can it use the existing LORA from SD 1.5 or LORA from SDXL?

@settima_ai from what I've heard it's a different architecture, therefore no

Im using that but then i get: RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)

@ppkkmoon @ppkkmoon请问是1.5的lora吗？我在这个网站里找不到太多的xl的hentai Lora

就比如我之前说的哥布林轮奸的那个没用xl的lora

pony系的lora和其它的不互通

chenkin他们用的底模一般是a3或者koha分下来的，被分到xl系里面了

pony及其分下来的模型都被划分到pony系里面了

（挺烦的）

pony和A3虽然都是XL模型，但是血缘关系差得很远，训练出来的lora都不能互通， 这点很奇怪……

来试试这个，后续我还会更新其他角色的~https://civitai.com/models/331888?modelVersionId=373825

its vae is broken（or the author does not have built-in），you need to use an xl vae alone

可选1楼的vae，或者c站内就有的xlVAEC_f1，直接用自带的baked_vae就是发灰的；顺便，如果你做放大，记得在encoding那一步连原本的vae，否则成图就会变得越来越亮

try offical sdxl-vae

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#break-keyword

@akerue_ thanks

@akerue_ but why use it? it just says it splits the chunk

@Darkhitman in case you forget to add comma in the prompt, i guess?

SD 1.5 has nothing to do with this. This is a PonyXL based checkpoint model. PonyXL is based on SDXL but is really it's own thing in regards to prompting which also means it has its own set of LoRAs.

@windir19 Is the pony model the XL model? I know that there are currently sd1.5 and XL. I just learned about pony recently. Is this a new XL?

@luouete PonyXL V6 is a fine tune of SDXL to the point that it is basically its own thing. The big difference is that it uses a quality rating system of points meaning use of scores in the positive prompt like "score_9, score_8_up, score_7_up, score_6_up" rather than saying "masterpiece", "high quality", etc. Also the dataset the images PonyXL and checkpoints/LoRAs are trained on use booru tags (tags used for image board sites like Danbooru) meaning you will get better results with prompts when using those tags. But again it is still SDXL so prompting doesn't have to adhere strictly to booru tags.

This model has no VAE baked in.

Since Fooocus doesn't have a VAE setting, you need to bake the VAE in webui before using it in Fooocus.

me too did you find how to fix it

@aliasolol I did!! You need to use a VAE. I use the sdxl vae: https://civitai.com/api/download/models/290640?type=VAE

find and use sdxl 0.9 vae

what needs updating?

@Hapse I think he wants a new model merge with more updates and refined models