Sorry, I might not be understanding your question. :) Are you wanting to know why I listed Noise Offset first or are you asking when Noise Offset should be merged into a model? If it's the first one, it was just listed randomly. And if it's the second one I'm not really sure. I used the extension 'SuperMerger' to merge this model. And every iteration I would test the model...if it started to lose the darkness, I'd remerge Noise Offset back in. So, I think it just depends on how many times you're merging something.

And if that doesn't answer your question, could you elaborate a bit about what you mean? :)

@woobly Glad someone knows! I was just pressing buttons essentially, face roll on the keyboard, get lucky results :]

@woobly Ooh interesting. If you're on CivitAI discord message me so I can learn the mystique of what merge settings actually do what. I think I picked new ones constantly, add differences... multipliers, etc ...and I hope you like the model! :)

Hey there! There's no such things as noob questions! :] Style-Paint is a textual inversion that I haven't released yet. I'll try to release it soon!

@vortex71 https://civitai.com/models/18052/style-paint-magic It's released now :)

When you say no vae, do you want me to add a baked in vae version instead then? :)

@ritcher1 my model doesn't have a baked in vae. I do have a vae available to download though to use with the model. :) but you can use whatever vae you like. I'll look into pruning it when I'm able to get on the computer. :)

@ritcher1 Just added a pruned version. :)

Just added a pruned version. :)

Hi there! It goes into Stable-Diffusion-webui - Models - VAE folder. :)

@beonfuct371 You're very welcome! I hope you got it working! 😊

@kaibosh No, if you put it into Models -> Vae Folder, then go into your settings it should just look like this https://i.imgur.com/e969mAq.png

@kaibosh Just make sure that it's activated like that in your settings for the VAE, I put a screenshot link in the other comment.  

@kaibosh If you want to do that, go ahead. I just use this VAE with every single model that I use, regardless if other models come with a VAE. So, I never personally use auto.

@kaibosh You're welcome, it makes me very happy to hear that you like it! I hope you're able to create some beautiful stuff! :)

Literally just did a few mins ago. :)

They are safetensors, not ckpts. Is there a reason why you'd rather have a ckpt than safetensor file version? :) Or if you don't mean that, if you click the arrow to the right of the blue download button, you'll see the pruned version able to be downloaded there.

@Unicom I see. I'll fix it later today and message you here

@Unicom I went ahead and added a pruned .ckpt (pickletensor) for you. :)

Thank you so much! I'm just glad that people are actually taking the time to read it! Really glad I could help out! Have fun creating beautiful stuff! :]

Hi there, I don't really understand this question as I inpaint with it just fine all the time. :)

@SyntheticSunsets No problem! ^^

I am regenerating/making new display images for prompts that use textual inversions and I am also going to try to make it more clear on how to better generate high-quality images. I have gotten tired of having to help/answer non-stop questions, I simply just do not have the time to handhold everyone through getting textual inversions, etc to work. So, I'm just trying to make it easier on myself. Although the display images do suffer now for a lack of flair without the textual inversions, but ¯\_(ツ)_/¯

@Sirtwenty7 You're correct! There are a lot of things that are not common knowledge with SD, and it's honestly a very learn-as-you-go type of thing! I try my best to explain things to people, I went ahead and wrote a guide about preventing double heads and the best way to upscale images...I also responded to a comment writing a detailed explanation as to why people cannot reproduce 1:1 images. And now I am also going through every single one of my display images and giving the generation data that they had before being upscaled...as upscaling changes metadata!!!

So, I try my best to help out and explain everything. I don't know everything though, so I can only go so far to share the knowledge that I do have. The only frustrating thing is when you try so hard to help someone, you can go blue in the face trying to explain, but still, some people just do not understand. I've had a lot of people get upset with me on discord...as I am open to trying to troubleshoot with people on discord in DMs. And when they find out that they'll never be able to 100% exactly replicate an image 1:1...they get extremely distressed over it and feel like model creators are somehow duping them.

But, I am now trying my best to alleviate this by updating guides and giving unaltered generation parameters that will hopefully help and also inform people as to why people get different results when generating images.

In any type of environment, be it SD, playing games, or anything...a lot of people forget what it's like to be new to it. It can sometimes be a very unwelcoming situation to step in, when you try to receive help or get your questions answered and no one responds, or if they do it's in an extremely condescending way. So, it's always important to keep this in mind, as sometimes people are truly ignorant and just do not know things.

But once you explain the reasonings, and then they're still upset, well then that's what kinda frustrates me and gets me down. But I am now trying my best like I said, to alleviate any frustration people have regarding generating things.

Regarding your question though, when you say class images...do you mean person/style, etc? Or do you mean the regularization images that people sometimes use?

There are A LOT of reasons why, you will likely NEVER be able to reproduce an image 1:1.

So, first of all, though, thanks for reading the guide on getting rid of the doubling up of heads/bodies, etc, I'm glad to learn that it's helped you now. :) So, starting from the guide, you know how I explained to get even further higher-resolution images you then run it through img2img with the SD Upscale script? Well doing that, changes the starting seed and then obviously a lot more settings. And because the metadata is recorded at the end of the image, it doesn't include what the image also started at.

Award winning, masterpiece, (sharp focus:1.2) beautiful mysterious mother nature( sitting against a fruit tree:1.1), (shimmering gossamer earth tones intricate woven gown made of flowers:1.2), surrounded by (citrus fruit:1.1), deep rich dark shadows, (up close from the side:1.1), earth mother, smiling, enchanting whimsical fantasy forest background with fruit laying everywhere, happiness, enchanting atmosphere, (cat eye makeup:0.85), (large eyes:1.0), enchanted forest background, looking at viewer, (freckles:1.1), cleavage, 8k, (high quality:1.2), depth of field, bokeh, 4K, HDR by (James C. Christensen:1.2|Jeremy Lipking:1.1)

Negative prompt: worst quality, ugly, deformed, mutated, easynegative, ((nude)), low quality, (saturated), low resolution, bad quality, b&w, bad anatomy, bad photo, bad art, bad proportions, (((mutated hands and fingers))), 3d, cartoon, (body hair), (leg hair), anime, (facial hair), watermark, logo, clothes writing

Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 9, Seed: 1105675196, Size: 512x704, Model hash: 46d105afa7, Model: FaeTastic, Denoising strength: 0.45, ENSD: 31337, Hires upscale: 1.5, Hires upscaler: ESRGAN_4x

Here, for example, is the starting information for the image that you're referencing. As you can see, it has different seed and generation parameters, because I then went and scaled it up via img2img. If you use this seed and generation parameters you should be able to get closer. But, I use Xformers, I have an EVGA 3090TI as you can see I have the ENSD set at 31337. So, it's highly likely that you still will not receive a 1:1 image as xformers and different graphics cards will change things.

Stable diffusion is a stochastic model, which means that even with the same initial seed, the diffusion process can take different paths due to the random noise added at each step. This randomness can result in small differences in the final image, even if the general settings are the same

Let me show you an example that I just did very quickly, I went to DreamShaper and I copied the generation data exactly for Lykon's starting image and I received this image: https://i.imgur.com/LRxBATx.png while here is his image: https://civitai.com/gallery/265136?modelId=4384&modelVersionId=24365&infinite=false&returnUrl=%2Fmodels%2F4384%2Fdreamshaper ...so, you see she's different. But what's important though, is that even if she is different...she still has the same tone and aesthetic look as the example image given.

Different hardware configurations: Stable diffusion can be computationally expensive, and the hardware used to train or generate the images can affect the results. Different hardware setups can result in different precision or rounding errors that can affect the outcome of the diffusion process...so like I said before, every computer generates things differently. You will never be able to get a 1:1 -exact- replica of the starting image. For all of the plethora of reasons I listed above...but to summarize.

1. If Upscaled via a script, the seed and information will change from the starting image, as metadata only keeps so much information.

2. Xformers changes things slightly.

3. Different graphics cards and different computer setups, I have even heard that different versions of A1111 generate things differently. I haven't updated A1111 in probably a month. I know a lot of people don't update often, as it regularly breaks extensions and scripts that people enjoy using.

4. Different settings in A1111, for example, Clip Skips, ENSD seeds, there is an option in A1111 that I also have checked, https://i.imgur.com/kTgZii8.png it changes things too.

5. Things like even simply copying the generation parameters wrong, if you accidentally remove a comma, a space...that is simple enough to change something.

6. Textual inversions, LORAs, you could also be using a different FaeTastic than me. All images were generated with the non-pruned 5g size safe tensor. Yes, pruned models also generate things differently than non-pruned models. I didn't have any Textual Inversions included in these outputs anymore, but if you go and try to reproduce someone else's example..it's very possible that they might.

7. Control Net...I did not use Control Net to pose any of the above display images. But I know that a lot of people do use Control Net to display images.

So, to summarize it can be a crap ton of things as to why you cannot reproduce an image 1:1. It's important to not get hung up on the fact that it cannot reproduce the image exactly. If the image comes out with the same tone and aesthetic, which I would say your image did, then that's a win. You can see the model is working correctly along with the prompt. And if you want your image to have a clearer face, then you need to do the upscale thing I talked about in the guide. And then even then, sometimes upscaling doesn't fix the face 100%...so then you need to inpaint the face to get it fixed. And then inpainting should always fix it if you are using the correct settings.

Anyway, I hope this helps you understand...and I also hope some more people read this. As this is probably the question I get asked the most. :]

@Whitevs Yep, sounds like you need VAE. :)

@pope_phred That's fine :)

You can use whatever one you want. There's no huge difference, other than size. People like non pruned models for training/merging. If you're just wanting to generate and save space, just go with the pruned. :)