@SG_161222 Thanks for the quick reply. Love your model and appreciate your work. :)

@SG_161222 So, how about diffusion_pytorch_model.bin should we load it in Automatic1111? It mentioned in huggingface.co

@SG_161222 Thank you very much, I am going to have a try, So the diffusion_pytorch_model.bin DOES NOT requird? Because I searched in my Automatic 1111 folders, I suspected I made some mistake during installation.

@leiyangcl781126 He was saying that the vae-ft-mse-840000 (the updated link) is used INSTEAD of the diffusion_pytorch_model.bin. :)

@SG_161222

@SG_161222

@SG_161222 thanks for the clarification!

@SG_161222 what if i use webUI?

@SG_161222 Appreciated it!

@SG_161222 thanks friend!

@SG_161222 thanks!

Yeah I've gotten the same error on all the versions, not sure why.

same error

same issue 4090rtx

same here also got this on Aly's mix as well, mine won't even generate a image

same error

Same error, version 1.3 works fine.

Same here

I too am unable to generate any images with this model. Im using Draw Things app on mac and it starts but then almost immediately stops generating.

Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion

Apply Settings and Reload UI

@ovalshrimp No joy.

@ovalshrimp did not work

seems its just a download issue from civitai , will redownload and try again

From the mod at Civitai on discord

"We just resolved an issue affecting files downloaded in the last 4-5 hours. If you're experiencing issues with things you've downloaded, please download them again now. To the individuals that jumped in to bring this to our attention, thank you!"

That hash matches realisticVisionV13_v13VAEIncluded.safetensors, which I have a copy of, but it appears to have been removed from the available downloads. I'm not sure why.

For a while there were a series of checkpoints that included the VAE. With 2.0's release, now I'm just using it with the VAE configured in Automatic1111 again.

I found the embedded VAE convenient, because I didn't have to enable/disable the VAE in settings for different checkpoints, remember which ones needed it, etc.

Either way, you'll be able to get the same results as this checkpoint if you use Realistic Vision 1.3 with the VAE configured.

@SG_161222 Thanks. I'll keep an eye out for the upload. The files at huggingface don't have the matching hash.

@Hexus I've been using the  vae-ft-mse-840000-ema-pruned.ckpt VAE and they're still way way off. Hm.

@SG_161222 hi! i'm using realisticVisionV20_v20.safetensors

@SG_161222 no sir, i downloaded the same model from the link that you provided and its working like a champ, thanks a lot for support and the model

What problem do you have with black models

Yes! This is my go to checkpoint as I get the best results using it but the skin is too perfect, it's like there's Instagram filter applied.

I've found that to be true of every model I've used, do you have recommendations for models that don't airbrush everything?

@SG_161222 Thanks!

you dont have to credit no one

re-download the model or switch to another model and then switch back

don't really know about this ckpt exactly, but try putting a big lens like 120mm or above, sharp focus, stuff like that in the prompt. Putting blurry ,DOF, depth of field, unsharp(?) in negative will probably help.

In photography, a large aperture number (= small aperture) usually creates a deeper focus. Try adding "f22, f16, focus stacking, small aperture" to positive prompts and "bokeh, depth of field, dof, narrow focus, blurry, out of focus" to negative prompt.

It seems the model has been trained on portrait with a blurry background. I tried with all the prompts, but no success so far.

me too, not success, seems portrait always has blurry bg

the words you put are represented in a "latent space" where similar things are next to each other. Its a bit complicated to understand, but luckily, its very intuitive to use. Just ask for what you want. big ol tiddies? "big breasts". in a field? "in a field". Best thing to do is read others prompts

@auserofthisname I think you're answering the question "how do SD models work" but what I asked was what role trigger words have for checkpoints as opposed to loras. Four trigger words are listed for Realistic Vision above, and I'm curious about whether that means those words do something special over and above normal prompting, or not.

@Onvisi trigger words are tokens the model has been specifically trained on. So this model might not respond to 'Nikon retro' as well as 'analog style' because the training class/instance captions specifically included 'analog style'. Being part of the model, they're less specific than LoRAs, but you can still weight the trigger words with parentheses or decimal values.

Im not the most knowledgeable on this but loras are essentially smaller trained models that you can add on top of prompt. They are topics ranging from specific people to art styles that the checkpoint you're using isn't trained one. You download them and add them to your webui model folder

LoRAs are simply mini-models that are trained on specific data. If you are interested in their training or simply AI stuff, I made tiny discord server focused on this kind of stuff. Let me know if you are interested

@IAmSteve what's your discord?

Loras are small networks that are trained on specific subsets or aspects of images. They work more or less like a normal stable diffusion model, but they're far more specific. They go in the models/lora folder in stable diffusion. You use them by adding their tag to your prompt.

Lora tags look like this: <lora:exact_name:1.0>, where exact_name is the name of the file for your Lora (excluding the file extension), and the number is how much you want it to affect the results. You can find plenty of examples of such tags on images here.

Automatic's webui also lets you click on Loras instead of typing them in. They show up when you click the Show/Hide Extra Networks button above the Styles drop-down.

@Syn_tzu Gamemaster2022#2506

Are you doing txt2img (vs img2img)? What dimensions are your images set at (512x512 or 512x768 etc)

@alephgates103 txt2img 768x1024.
I added (full body:1.6) or 1.8 and it seems to help a little.

I see. Thank you for your reply.

stable difusion

@SG_161222 Which one (VAE), safetensors or ckpt?

And what inpainting version used for?

@SG_161222 why i can't generate high quality pictures? it's looks low quality pictures & a bit blurred, any advices?

试试增改关键字

这个能拿来图生图吗？我局部重绘总是和原本的肤色不一样，urpm就没这个问题。。啥情况

I just had this happen for the first time with v1.4 of this model! I got a telephoto lens or a camera body in some of my gens. I used "Nikkor" and "Nikon" in my prompt. I used to always get Hiper, Hipper, etc as text when I used a friend's prompts with "Hyperrealistic". I always wondered if that meant it didn't understand it. But I've seen some weird typos like "3555mm". We never had telephotos THAT long in my day. 🤣 Nor negatives that large. I guess it's 35mm and 55mm squished together. I never see that on a shirt though. 🧐

I did extensive testing on detailed photography gens using a comic book character lora and as many "realistic" prompts as I had to make the character look real. Here's the ones that worked very well: high quality photograph, high skin detail.

I tested over a dozen HD prompts, and yep, just those two made the biggest difference. So aim "high"

@Electroverted Excellent! "high" makes much more sense than "hyper". Did your extensive testing happen to involve placement within the prompt? Using it up front vs. middle or at the end?

I've been experimenting with lens names in the prompt. "Petzval", "Helios 44-2" and "Holga" have a noticeable effect.

@parallelepipedon I'm sure placement makes a difference, but that would be a lot of testing! From what I hear, start and end prompts have more strength. My detail prompts are at the end.

Well, his model is the work of many different models.

Plus, some of the checkpoints have been trained with copyrighted material

Wait how do you know that?

the license information and what is allowed and what is not allowed is located in the bottom right under that download box area.

same issue and question. i am also hoping for someone to shed some "light" into this question! :)

Write in the prompt, for example: feet shod in strappy heels, etc.

More detail in your prompt can usually help include what you want. Adding "feet" to the prompt will improve the chances that they're included.

If part of the image is still cut off, you can use Outpainting to add it back in. With Automatic, you can set one of the img2img Outpainting scripts to extend in the direction that's missing.

OpenPose + ControlNet works every time, guaranteed. You can also use revAnimated to generate a full-body pose, then run it through img2img + this model.

set your size to 512/768 you will be more likely to get a fully body image.

add: ((full_body)) to your positive prompt.

add: ((out of frame)) to your negative prompt.

Change the resolution of the generated image. When I switched from 512x512 to 512x786 I started getting more full body framing. Also, include at the beginning "full body photo of xyz". I do batches of 8 and probably 80% will be full body.

@loafkeeper 512x786 is a good tip; thanks, it def helped generate images that are in portrait orientation!

And you never thought of adding the nationality you want as a positive? Example "An English Male" or "An African woman" .... I cannot believe you're moaning about this when it's pretty simple.

Forget it, I was too hasty, I'm using it and it's great. I apologize for the negative comment.

I don't ever prompt for nationality, and I am getting mostly caucasian.

Try using ControlNet with it using the appropriate input image. That should help.

I use jut basic txt to image, just needs a prompt https://civitai.com/images/697953?period=AllTime&periodMode=published&sort=Newest&view=categories&modelVersionId=29460&modelId=4201&postId=191289

Do you have any error message in your console? For instance this just happened to me when controlnet was enabled but no pic was present, there was an error pointing me to the problem in the console

Did you enabled the recommended VAE?

Because you are using Intel HD 4000 GPU bro

you probably have an embedding at 15 instead of 1.5. Then no matter what you do you onna get pixel or ... spaguettis ^_^

So, modify your prompts to remove this issue.

You are responsible for what is inputted into the model for creation. You need to be more specific.

girL means female chil, say woman or teen18+

1girl is a booru tag, its only relevant for anime models or models that imitate booru tags. In this one use "woman" if you want to generate an adult, typing in an age can help too if you want to be more specific. Check other images posted on this page to get an idea of how to prompt for a model like this.

@AiShawArt I'm not a newbie, I'm just testing this model. In my experience with other realistic models, the images entered into 1girl's production are almost always pictures of girls over the age of 16

@cactusman I know that typing other words or adding some negatives can generate mature girls, but I haven't encountered this in other models: six-year-old girls generated by the word 1girl

Well, you expect that outcome because the other models are biased. A GIRL is a child a WOMAN is an adult. Those words aren't even ambiguous. So a prompt for a girl creating an actual girl is more accurate, don't blame this model for creating a girl, blame the other models for being so incredibly (porn) biased that asking to generate a girl doesn't generate a child.

@90S and those realistic models mostly suck on generating different ages.

Try to use "child", "childish", "underage", "loli", etc... in the negative prompts. Do it literally all the time no matter what. It's pretty jarring when you forget and you get weird stuff in the outputs specially if you are doing NSFW stuff and you really don't want that to happen.

Also I've noticed "girl" gets young-looking people even if I specify something like 25 years old sometimes. So try to use the word woman instead. Not all models have this issue but these tips should help

Add fused bodies in negative prompt with a good weight like 1.3 or 1.4 and in parenthesis.

Model based on BooruDatasetTagManager, so no need space

Can you show an example ? i have not any problems with this model.

randomize the seeds

Tested with 2.4K images and 1024,1024 size, default kohya_ss settings and batch 2
I would say I managed to improve quality, but failed with persons (too much similar faces)
So I will replace some images fore sure..

@pastuh Thank you for the data help you provided, I used 512*512 size,other training data are commonly used,but the trained model generated pictures are very blurry. does a 1024*1024 size to make the effect better? But my computer doesn't seem to support me to validate my ideas，so ...(translated)

@SG_161222Thank you very much for your honest answer~I'm currently trying to learn stablediffusion and trying to train beautiful models, but the results are not satisfactory。I wondered if I was having a problem with the number of image sets I used for training, because I only used a hundred images and the outputs were very scary.I don't know who can answer my question, maybe a thousand pictures can solve it? I will continue to try. Thank you again for your answer!o(〃＾▽＾〃)o

@SG_161222 I am curious about how did you merge models... I am just a newbie here, so it would be great if you could also explain what exactly mearging models mean?😅

Same here... It would be v helpful if you could give some more details about the dataset of images you used and other parameters during the training...

While I can't answer your first question, the inpainting model is going to be the one you'll want to use when inpainting, (and maybe img2img? I do). So generate the image with the regular checkpoint, but if you are making any changes to an image, switch to the inpainting checkpoint. Use a model made for inpainting is the difference between "Inpainting doesn't work very well" and "Wow. This does exactly what I need it to... usually."

@Toto_in_Kansas thank you very much

@SG_161222 Thanks! 

i am kind a new here... so i ask very naif question. what is the VEA? pls this is a real qestion.

@emiraydin84489 A VAE is a file that is used along with a model/checkpoint. It can add more colour, contrast, vibrancy and so on to the final image. For example, if your image looks washed out then a VAE can fix it, It's definitely worth having a couple for when they're needed.

@SG_161222 please bake it in!!! ;)

Though compare vs Dreamlike

but yeah both are excellent tuned 8 o clock reconstruction models, with different paths

both are pretty excellent for everything POV related

Dreamlike is overall a tad more unique maybe from it's Genetical Base but its a integral part here maybe even its driving core

Would be very interesting to see what happens if Dreamlike is left away from the merge

Would it fall apart entirely ?

vae is neural network that transforms the input & sends it into the 'latent space' where it gets feeded into clip then gets mapped by the unet (still in the latent space) & then again receives the sampled input from the latent space, encodes it back and produces the result. Remember a model has 3 components vae, clip & unet.

thank you

how does that work with an app like Draw Things

keep changing the ethnicity and specify ages like - 24, 28, 32 year old

That's how SD models work. If you don't change your prompt, you'll get very similar looking people.

@redmagechris2916 not all if you use LORA the face will look different,, because LORA like DollLikeness uses the same dataset of photos of people, if you don't use LORA maybe it will help.