@SG_161222 Ok, but they don't do regular generating, as well? (thanks for the reply!)

@SG_161222 Ok, thank you kindly!

@SG_161222 Is it better to merge models with the non-pruned model, or does it work equally well with the 1.99GB models?

@SG_161222 Is it better to merge models with the non-pruned model, or does it work equally well with the 1.99GB models?

Yep there seems to be an issue with this version, cant do lora training on it, or merge models with it

could this be a problem with the pruned model only? don't want to DL the full model if it's also broken.

@jbmetafilter813 It happens with the four models (realisticVisionV60B1_v60B1), with vae and without vae, pruned and without pruning

@SG_161222 just trained a lora on this and it worked perfectly, better results even with it than 5.1 using the same dataset nothing changed but the model you provided in the google link. Using the latest Kohya_SS

T@SG_161222 Thanks, this one works

well probably this is not yet the full version of the model?

@gpaulelli  I hope

nah i suffered deformation also if i use my clothing lora. probably its still beta release, not full version of the model. so i'd rather stick with 5.1 for now.

This model is highly fine-tuned to work with enhanced resolutions (see HF page). As the other side, it may be rather different from other models in many aspects.

Getting exact same results. The model is aggressively adding women (especially pretty women) without any mention of a woman, explicitly saying a man, and then keywording every synonym for woman with heavy emphasis in the negative prompt. Even with doing all that, a batch of four will have two women subjects in the results.

i agreee, with the latest beta version is impossible to generate a male character

@eg1989 at this point I think it's better to stick with V5

The best prompt

@eg1989 good prompt

It means VAE is baked-in and you don't need one.

@infernahermit846 I was going to ask the same question actually, where would the model need to go? I"m using Fooocus, started playing around with that a few days ago, but I'm just a total noob, still looking for a beginner's guide :) I tried putting it in the VAE folder, but it doesn't show up in the pull down menu option.

@SG_161222 Thank you. That would be great. Because the previous versions of your model often give better results with celebs than using most of the high-rated Loras of the same celeb on civitai.

@ImAbbieKitten Well, don't count me out yet! (I'm not popular yet) but I think I have pretty good grasp on celebrities faces. I've seen and tried a lot of loras that didn't quite feel up to the task of generating an acceptable likeness of the celebrity. So I've started creating my own Loras. Check mine out in your free time. My current list is small but I believe it to be a rating of about 80% to 90% of the actual person's likeness. No pressure but I think its worth checking out. Oh, and @SG_161222 if you want, you can use my lora models to assist in your training or fine-tuning. You have my permission.

I will try, though it's not like I didn't try. Seems like negative prompt is close to non-existent for this model. Thanks for your effort anyway, v5.1 is my most used model.

There's a bias, but it's certainly not forced. Even including the most generic prompt tags for clothing like "clothed" or "outfit" give me non-nude every time. I've never found including nudity tags in the negative prompt to be all the effective in any model that tends toward nudity.

any 1.5 penis lora should work

try the titsonastick lora

I don't have any issues with appropriate prompting. Just put some tags like "busty girl, big boobs" in your prompt and add "skinny" or "very skinny" tags. Then in your negative prompt, something like "wide hips, curvy, thick thighs, plump, huge ass"

@cidolfas666 thanks, so when you say tags in prompt you just mean the words in the prompt box, e.g. (big tits, fake tits, skinny, toned legs)

@tcpstackmaster971 I have browsed all the fake tits Lora and they all heavily skew the image to their training images, usually bimbo and often have negative impacts on the quality of the image? I'm not sure on the impacts of Lora as I've never had good images, should I change steps or cfg? does it matter the base Checkpoint as long as it's SD1.5 (or correct for that Lora?)

sounds like a weighting issue.
pun intended.

@rogue8888992 yeah in the prompt box, by "tag" I just mean that they're separated from the rest of the text by commas.

Im so sorry for asking the stupid questions but how do you use this LoRA?, this is kinda the first i download and dont know how to make a correct prompt, im using fooocus

@rodrigocruzglez679 This is not a LoRA. This is a base model. However, I have never used fooocus, so I don't know how to plug a downloaded base model to it. Probably you should place the file into a certian folder. Don't hesitate to use Google for details!

@rodrigocruzglez679 just download it and put it in the "models" folder of whatever UI you are using

@SG_161222 thanks for the reply, it's a great model nonetheless, i'll eagerly wait for B2

Try adding ethnicity such as "Caucasian male" and negative prompt words for female only body parts :). "stubble" if male has any facial hair helps also.

it's true, doesn't accept men

RealVisXL V2.0 https://civitai.com/models/139562?modelVersionId=169921

I have the same issue, this version i not working well with character loras

It's definitely not just you... the best results with this model are top tier, but it takes a lot of prompt adjusting and sample quantity to get one.

I have same issue. With loras I get blurry images

That's it. I see. Thank you for your detailed answer.

For reference, there is a section in the wiki of A1111's webui detailing the process for generating inpainting models using this method: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/How-to-make-your-own-Inpainting-model

@c144 Thanks a lot, I will try it.

6 isn’t working on here?

Your prompting is matching tags with exposed areas. Try putting "nudity" and specific body parts in the negative prompts, "NSFW" isn't universally trained on realistic models. I don't train for it at all on mine.

Well, pretty much all models are mainly females, tits and pussy. That's what (sadly) stable diffusion lives on, on a community of degenerates. lol

@moedow326 Facts

@SG_161222 Thanks!

@SG_161222 One more thing: If I get the inpainting version, this means I can ONLY inpaint with it? Can't use Img2Img / Text2Img? Which version lets me do everything?

I managed to find an old reply from SG. So apparently, the inpaint version is better at inpaint but can generate, and the regular version is better at generating but can inpaint.

It's so difficult to pick checkpoints now. I've hundreds by now and under each I find at least one comment, talking about how great the model is. lol

I'll keep this one though, got to use it more to confirm it's great.

I can confirm that reatlistic V5.0 is currently the most realistic model. closely followed by absoluterealityv1.8

@Timmy_randall is there no point to get V6 then? or would i expect the same/slightly improved results than V5? i'm new to all of this so i don't know what really happens with new models unless the creator thinks they're improved in some way..?

@Narz I think it is a personal experience. In my case I have tried many models to get images of people in many poses where the most difficult challenge is to get a realistic skin texture in full body or half body poses. There are many models that make very good quality close-up portraits, but when the body moves away the problems appear. I have achieved the best results with Realistic Vision V5.0, but maybe if I had spent a lot of time with other models trying many settings, testing with Samplers, different lighting instructions or skin textures at the prompt ... it's a lot of things, and you don't always have time to dedicate to do a lot of tests. It's complicated. Look on civitae for a model that you like at first sight, look at the examples and try to get the best results with the simplest possible settings and prompts, and if you like the results stop looking at other models.

@Narz This is just a beta version of v6.0, that's what the B in the version stands for. Beta version 2 would include more males in the training. The author confirmed this somewhere else in the comments.

I assume V6 will be improvement once it's finished, as i'm getting amazing results with v5.0

same!

I join the question, we are waiting for the author's answer

@SG_161222 期待大佬的模型~~大佬加油！

@SG_161222 感谢！

@SG_161222 thank you!

@SG_161222 for the time being what would you say is the perfect model to try right now and also can you like tell all the prompts words etc because all of that stuff is mostly behind paywall and people form poor countries like me that is bit much for words

yes please!

👍

My strategy whenever this happens (doesn't always work but worth a shot):

Go into the Checkpoint merger in A1111, select the target model and the base SD1.5 checkpoints. Choose a weight of 0 for SD1.5 and 1.0 for your target model and have it merge. Then convert the newly generated checkpoint to diffusers.

https://huggingface.co/SG161222/Realistic_Vision_V6.0_B1_noVAE (official)
https://huggingface.co/NickKolok/Realistic_Vision_V60_B1-conv (reconverted by me, may be a slight more lossy)

Im getting the same problem.

They're different models so unfortunately they're going to behave differently. You could try merging V51 and V60 together to get something closer to your V51 results though.

@aicatgirls3d. Thank you for your advice. How can I merge 2 models? 

I would have to agree.

yup! gotta agree

Same

Every model has it's own sharing policies, but don't worry, most of them can be used for commercial and non-commercial purposes. I would suggest checking the description of every model you use, even if you don't understand the jargon, you'll still be able to understand what kind of use the model permits.

Same problem here

look up tiling.
https://www.youtube.com/watch?v=hZ_qtn5SHso

@ThereDX Thank you very much for your support

Yeah, go ahead and blame others for your lack of knowledge and understanding. Totally clueless.

download the realisticVision 5.1 checkpoint file. Put it in the models>Stable Diffusion folder. Change to the new model in the drop down on the WebUI. Use the template given in the notes on this page to get good photorealism for human subjects.

It's baked in

Already inside the model, you can use without external or other VAE, just set to none.

Yeah 5 is way better specially when you use IPAdapter

It's not done yet

v4 is good, v5.1 is better.

v5.0 is bad, v6 is worse.

@dillon101 When it will be done ?

@codegix not my model, ask the creator

The inpainting model helps better quality when you use it in A1111 inpainting tab mostly generate new things and edit things

@mridiot918497 ok. i got it. tks u

Welcome to the world of image generation

The first is civitai's doing as the title of this post is realisticVisionV60B1, one of the model files is pruned i believe and hence the halving in size

@SG_161222 well thanks for the update! can't wait to try it!

@SG_161222 That sounds promising. Now I'm excited to try it as well.

Any chance of publishing any explanation or anything in any written form that provides any details relating to how you ended up finding the best trainign parameters you mention and/or anything providing any details about any specific B1 problems that were identified and addressed in B2, unless the problems just went away simply as a result of achieving a better model fit?

I've been trying to train modles, both using dreambooth and lora/locon methods and based on my completely anecdotal subjective experience, it's seemed like getting a good result has been rather more challenging compared w/ my experience using other base models in the past (though they were of a style that was slightly less photorealistic), however, in the few instances in which I did manage to get it to work well for me using this as the base model, in those cases, it worked very well, actually giving me the best results that I have achieved thus far to date.

So any informatioin related to model training, especially anything related to this model in particular, could end up being extremely useful in my future attempts at finetuning dreambooth/loras starting from this as the base model.

@SG_161222 Where can I find the update stats? Apologies, fairly new with all this.

It's generally considered the better version

@dillon101 I doubt that's the case here, this kind of reversion has been occurring for a bunch of different models recently and more than one of the model creators weren't aware until I pointed it out

Are you using a fix seed ?

Their output should be about the same, but the full 4GB model is better if you’re going to use it for a merge.

did u find a fix?

@tnythebadasstiger914 So far, no.

same issue

Oh that's interesting. I found that 5.1 looks more realistic, at the cost of ignoring details of my prompt, whereas 6.0 beta is more literal when it comes to interpreting my prompt, at the cost of body mutations and melted fingers.

@skyeve yeah i can say the same, you get good results but horrible fingers and mutations at times, I Always create a batch of 10+ images so I will end up with at least a reasonable few...

Not if you prompt it right. I use this model for use in Pregmod for Free Cities, and it helps generate the positive and negative prompts for each Stable Diffusion image generation. Try this prompt for V5.1 of this model:

Positive:
full body, on bed, cinematic photo, highly detailed, professional, gritty, sharp focus, high budget hollywood movie, dynamic cinematic lighting, color graded, (indistinguishable from reality:1.2), realistic RAW photo, 4k, 8k, HDR, (depth of field:1), bokeh, ultra detailed, bloom, ray tracing, (sharp focus:1.2), intricate high detailed, intricate details, hyperdetailed, absurdres, professional, award winning masterpiece, Fujifilm XT3, fair skin, caucasian, girl, teen, teenager, 18 year old, best quality, standing, arms behind back, slim, (completely nude:1.1), pussy, nipples, small breasts, flat chest, neat hair, long brown hair, green eyes, thick eyebrows, Russian, (smile, looking at viewer, confident), earrings

Negative:
back, (standing:1.4), (worst quality, low quality, normal quality, lowres, low details, oversaturated, undersaturated, overexposed, underexposed, grayscale, b&w, bad photo, bad photography, bad art:1.1), (watermark, signature, text font, username, error, logo, words, letters, digits, autograph, trademark, name:1.1), (blur, blurry, grainy), morbid, asymmetrical, mutated malformed, mutilated, poorly lit, bad shadow, draft, cropped, out of frame, cut off, censored, jpeg artifacts, out of focus, glitch, duplicate, (airbrushed, cartoon, anime, semi-realistic, cgi, render, blender, digital art, manga, amateur:1.3), (3D, 3D Game, 3D Game Scene, 3D Character:1.1), (bad hands, bad anatomy, bad body, bad arms, bad legs, deformities:1.3), bad anatomy, deformed, (mutated body parts), (bad quality), seams, text, signature, watermark, distortion, lowres, error, missing fingers, extra digit, fewer digits, mutation, extra limbs, disfigured, out of frame, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, mutated hands, fused fingers, too many fingers, long neck, multiple heads, monochrome, bad feet, unclear eyes, low quality body parts, missing body parts, disproportional body parts, indistinguishable body parts, branched body parts, rheumatism fingers, melted fingers, less details, dark skin, asian, beard, mustache, boy, man, elderly, adult, 30 year old, 40 year old, low quality, dark skin, asian, beard, mustache, boy, man, elderly, adult, 30 year old, 40 year old, low quality, clothes, jeans, underwear, pants, shorts, skirt, panties, angry, looking away, tattoo, sick, ill, pubic hair

DPM++ 2M Karras

CFG: 8

Sampling Steps: 30

Resolution Height: 768

Resolution Width: 512

ADetailer for the face with default settings.

You'll have to excuse some of the contradictory terms in both the positive and negative prompt. Certain default parts of the prompt cannot yet be customized in Pregmod yet so instead I just add the term I want to remove from the positive prompt into the negative prompt but at a hight weight, like (standing:1.4).

Just what I was looking for, thanks for the heads up

this is a checkpoint... Maybe read the file name

AFAIK it just means a VAE is included in the safetensors file. So this overrides whatever VAE you have set in your GUI.

@LurkingAI I was looking for this info exactly, thank you!

This is a SD1.5 model, the reason you probably had worse performance on other models was likely because they were SDXL models. SDXL models require 1024x1024 resolution for the best results, as opposed to 512x512 for SD1.5, they're also larger and therefore slower to run, even at the same resolutions.

Any regular SD1.5 model will run at the same speed as this one, you can filter CivitAI for specifically SD1.5 models.

@mylo1337 cool