It means to rename the yaml to match the safetensor, so it's like yiffymix4.yaml to go with the yiffymix4.safetensors

@PsySpy You commented nothing pungent

@PsySpy That still doesn't explain anything though... if two of the same file exits in the same folder then it's just gonna be renumbered

Did you set cfg rescale to 0.7 ? Its a vpred model so you need the config file in the model folder and you need to set cfg rescale, typically to 0.5 or 0.7.

@morric58 Thank you, I added the config file, which made pictures apprear. I don't have a CFG Rescaler tho (it seems to be an additional module for SD and the config file doesn't have any values with this name) and the installation seems to be more bother than worth, so I'll leave it at that and use a different model.

@EternalTrash Read the 'about this version' text at the top of the page, just below the download options for instructions. You need the yaml config file and probably want the a1111 extension linked, or click the link to 'easyfluff' in the expanded readme and you can get comfyui module from its model card instructions.

I want to know as well, same issue!

read the about this version description and follow along, works fine now!

Weird how 4.0 seems to work just fine on Easy Diffusion but, even after following all the directions under 'About model', I still can't get it work work in webui AT ALL. Oh well

Thanks :), it works now. I just have to tinker with it a bit to get the optimal output. And while I have your attention, I just want to say how much I and the rest of the community appreciate your models and the work you put into them. Keep up the good work, you are making a difference <3

"This checkpoint includes a config file, download and place it along side the checkpoint."

@Mobbun I did it, but still don't work, the other versions works finely, but this one, idk why it don't work,

"This checkpoint includes a config file, download and place it along side the checkpoint."

@Mobbun Did this, made sure I had CFG rescale, and it comes out looking quite bad using the example prompts and settings. Will just have to stick with v36 until some kind of tutorial comes out.

But... My lora training doesn't work! I tried this for kohya trainer but my lora models image are blurry.

"Check if you have the CFG Rescale extension enabled, if not, enable it - if you don't have it in your Extension list, add it using url https://github.com/Seshelle/CFG_Rescale_webui"

You say this, but how does one actually install it?

Figured out how to install with URL, but no idea what a VAE is. Images are coming out looking quite jank.

Cheese and crackers. Okay, so you weren't joking about needing that VAE. My google-fu is weak looking up install instructions for this stuff. Cheers.

@RetroSharkBite look buddy, I edit for kohya trainer and not stable diffusion webui.

I could try but how?

@RetroSharkBite @mobbun Haha! Now I know what that was! Now they are not blurry for my Lora pictures and everything runs normally. what that was, that was on "v_parameterization". I didn't know that I could set it up, but anyway, I've already set it up and it's running normally again! :)

I hope!

YES!!! https://i.imgur.com/VtPpGX1.png That worked really well and of course I'm happy. :D ;3

@mtndogs895 I'm sorry, what? That's supposed to be a joke, right.....

I'd check the year. That's from 2022, over a year old. Funnily enough, you're not the first person I've seen to make that mistake.

My bad. Thanks for pointing that out.

Okay, figured it out:
You need this ->https://civitai.com/models/118561/anything-kl-f8-anime2-vae-ft-mse-840000-ema-pruned-blessed-clearvae-fp16cleaned
Put it in your VAE folder. What's that? It's in your models folder.
In the Web UI go to setting > User interface.
Under [Quicksettings list] click the drop down arrow and select 'sd_vae', scroll up and press apply, then reload. Now next to your model selection select fancy new VAE.

Tada! No more deepfried images with v40

@RetroSharkBite Worked like a charm! Thanks a lot!
here is the image i got: https://puu.sh/JXBtU/b9e1d03895.png

Same deal with what I'm using, they should add that to the usage instructions.

But the instructions say you're supposed to use the vae kl-f8-anime2
edit: also just noticed that the vae isnt even in the downloads anymore in this version...

@rantas I can't say for certain, but I don't think that part of the instructions was there when I posted that or when I downloaded it initially. Either way it's there now, which is good.

Download the VAE

@chilon249 Thank you for replying. So, I am not doing anything wrong and it's normal that the images are darker and stick more closely to the prompt without adding much detail freely, resulting in more "dull and empty" backgrounds?

The backgrounds remind me of fluffyrock now... Are you sure CFG Rescale isn't required?

Gotta use the config file.

In other words - you were relying on model artifacting to get randomness and vpred has exposed the flaw in that practice?

Not sure what you're talking about with style homogenization or greyscale output though. You're not using a terrible, giant negative are you?

@nunyabusiness There is a certain brown tint that's apparent on EasyFluff and likely yiffy4.0. It's due to the fact that it's a vpred model merged with eps. I don't think pure vpred models have these problems. I'm not sure how 4.0 was merged, but it may have exacerbated the brownout.

Ive come to the same conclusions. The washed-out palette of clothing was especially noticeable. v40 is not an improvement, nor a step in a different direction, but a step off a cliff.

The point of an AI generator is to work alongside a random seed to generate a variety of images. If you plug in 'running' as a pose and aren't getting a variety of running poses that's as broad as reasonably expected from the training data, that is an issue. Especially when the tagging of the training data isn't robust enough to deliberately summon up the edge-case poses.

I'm also not sure what to think of v40.

People said yiffymix doesn't follow the prompt accurately and is too random. But I don't actually mind. I like using a lot of wildcards for backgrounds, poses etc and have yiffymix be "creative" and flesh it all out with details that fit into the picture. It's amazing what kind of detailed and coherent environments you get just by prompting "jules verne" for example. Always fun to make a few hundreds of gens at a time and then see what the model came up with while browsing through the batch. And the results were always bright, saturated, detailed and varied, the faces and bodies tended to be beautiful, even without any artist tags.

Now everything looks very similar and overly dark. Backgrounds are simple or missing entirely. Some tags that didn't seem to do anything at all before seem to be more effective now. That's true.

I don't want to pass a final judgement here, but at this point I don't really know why I shouldn't use fluffyrock instead. They seem more or less the same now.

Also, I've gone through the various tags and think I've found that the tag species I'm using (panthera) is excessively dark looking from the front, and is primarily rim-lit, and top-lit, which washes out pretty much any image since I can't seem to get a source of light from the front for illuniation

Have you tried the vae in the description? That mostly fixed it for me

@loppirr yes, it did help, and I found using CFG Remix made a major improvement as well!

Have you followed the (required) instructions for running this model? Especially needed for the latest v40, where you need extra stuff since it's a v-pred model.

A lot of people, including myself, are getting amazing results, so maybe check your setup.

Can confirm, added that between Load Checkpoint and sampler, Just Works (tm)

same

@widic47081 @solvenn7796 Have you both read and properly followed the instructions and recommendations of this version before claiming it "doesn't work"?

v40 is a v-pred model, you need the config file otherwise you'll get noise. CFG Rescale is also pretty useful.

I'm also having issues. I placed the 'config file' next to the checkpoint? I think?

This is an issue for many, many models. Try using the regional prompter extension if you're making an image with two distinct characters

@C1yde To add to that suggestion, also try defining your characters in separate areas of the prompt. I find if I define multiple characters at different places, it more easily separates them instead of merging the concepts together.

@sansmia Could you give some prompting examples?

@C1yde @sansmia This wouldn't be an issue, but if you want the characters to hug or kiss. This becomes an issue. I've tried telling it Right and Left. Front and Back. Even tried controlnet to get more control over the models. The only thing that seems to work is controlnet and then make your character in the area using masking. Then make the other by masking out another area. It would be awesome if it could just differentiate between the fact that there are two different characters. But once it sees one Lora - addled (addled like a sickness) prompt.. That the end of it. It stops listening and now I have 10 Nick Wildes or 10 Judy Hopps. This is the only reason I can find that Loras just aren't great. They allow someone to use 1/3 of the computing power to make, and less pictures, but then you get a partial model that is good at one thing and doesn't like sharing. :D I'll test out both of your suggestions, because it seems interesting. I would love to actually be able to tame some of these crazy Lora addled checkpoints.

It seems to work fine using TensorRT for me in version v40.

@yosh599 Yep me too, tensorrt is just suuuuper picky

It might help if you provide the prompt or even better the complete meta data (like seed etc.) Without it's hard to give some helpful advice besides the usual "put nude, nsfw, etc." in the negative prompt...

What's your computer's GPU? 25 minutes sounds like it's running on CPU.

@PapaPapa He has a Mac Mini so he has no gpu outside of the integrated one Apple built into their M2 processor. Macs are a pretty bad time for this sort of thing unfortunately. The GPU is the powerhouse of the whole thing, and without one, crunching on a CPU takes forever.

It might be a VAE issue, or maybe you need to add the CFG rescale extension.

@PapaPapa How do I add the CFG Rescale Extension?

@Steeltron2000 It should be under the extensions on A1111. https://github.com/Seshelle/CFG_Rescale_webui

That one

@Steeltron2000 The extension can be installed from the Extensions tab by copying the repository link into the Install from URL section. A CFG Rescale value of 0.7 is recommended by the creator of the extension themself. The CFG Rescale slider will be below your generation parameters and above the scripts section when installed.

@PapaPapa Ok How do I download it?

either can be negative embedding issue, like badyiffymix applies red tone almost everywhere, or boring_e621 applies yellow-green to many styles

@silkytail113 Ok

It's an issue with merging vpred and eps models. You get brownout and the yellow tint. It's also apparent on Easyfluff, though it's less severe there.