@epinikion Thanks for the clarity! These are amazing models and I, too, like to use a variety. NaturalSin and Epicrealism and Epicphotogasm are my go to :D

@epinikioni i found some remnants of a logo in 2 generated pics (both with a mountain view) using your model. thought that might be from a picture in your dataset.

@epinikion I'm using automatic1111. I keep trying but it just puts me back on version 5. I'm going to install this on another PC and see if I can re-recreate the problem.

@epinikion thank you for confirming this for me and thank you again for your work! I use ER5 almost exclusively. =)

UDPATE: So, look, this is a problem I have on my machine and others had this with other models and other Automatic1111 updates and I found a solution on Reddit from six months ago. Basically, this guy said that if you use the OpenOutpaint extension and then load the model from that tab it will work in the text2img tab. So for your edification @epinikion and for anyone else with this issue:
1. Go to Extensions tab.
2. Click "Available Tab"
3. Click "Load From:" button after ensuring repo is "https://raw.githubusercontent.com/AUTOMATIC1111/stable-diffusion-webui-extensions/master/index.json".
4. Search "OpenOutPaint" and install.
5. Hit "Apply and restart UI" -- if extensions are loading forever, go into "Settings" and hit "Reload UI".
6. Go to OpenOutpaint (I did not need to restart with --api) and then change the model. It doesn't throw the errors like in text2img and loads.
7. Go to "txt2img" and run. I confirmed working on my install! =)

Thank you for all your work and I hope my solution helps someone. Thank you to the community as well.

See my posted pics, trick is to not use certain keywords for details related to women, see prompts on my image post to get an idea. Is a pain though if you don't know.

Spot on. Uninstalling this model. This is meant to be some blokes wank engine.

Nope. I never make waifus. Read my other comment. I just did an excellent old man. But I did use I2I, because it was a particular old man and a bad photo reference. And not enough other photos to train a good LoRA, I don't think. You are using 1girl, girl, lady, woman, breasts, boobs, tits, etc in your negative, right? I'm surprised how many people who don't do any anime-related stuff who don't know "1girl" is big keyword. And so many of these mixes are mixed with some model that mixed with something that goes back to NAI. Those buggers were rigorous, and used a standard.

You do not need inpainting models for inpainting, i always exclusively inpaint using non-inpaint models and never had any trouble with it, just make sure you use ControlNet with a good Depth map that clearly show the form of the subject (breasts, butt, belly, etc...), so the AI can properly understand it.

@SweetHentaiAI no we actually need inpainting models. default models are very bad at inpainting real life pictures and they are only good for images that are ai generated by themselves

@SweetHentaiAI Default models are bad at inpainting. It's been proven many times before. Inpainting models are more precise for the inpainting tasks.

I feel your struggle. With this model and most female biased ones you can get around it if you use a negative like:

(Long hair, 1girl, Woman, feminine, breasts, girl, lady, female:1.4),

Or, if u haven't seen it yet, VirileReality's new beta is 🔥🔥🔥🔥

wrong, i can generate everything with it, but is super realistic to generate peoples

out of memory i guess, try smaller resolution perhaps

Error occurred when executing KSamplerAdvanced: mat1 and mat2 shapes cannot be multiplied (2464x2048 and 768x320) File "P:\ComfyUI\ComfyUI\execution.py", line 152, in recursive_execute output_data, output_ui = get_output_data(obj, input_data_all) File "P:\ComfyUI\ComfyUI\execution.py", line 82, in get_output_data return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True) File "P:\ComfyUI\ComfyUI\execution.py", line 75, in map_node_over_list results.append(getattr(obj, func)(**slice_dict(input_data_all, i))) File "P:\ComfyUI\ComfyUI\nodes.py", line 1270, in sample return common_ksampler(model, noise_seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=denoise, disable_noise=disable_noise, start_step=start_at_step, last_step=end_at_step, force_full_denoise=force_full_denoise) File "P:\ComfyUI\ComfyUI\nodes.py", line 1206, in common_ksampler samples = comfy.sample.sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, File "P:\ComfyUI\ComfyUI\comfy\sample.py", line 93, in sample samples = sampler.sample(noise, positive_copy, negative_copy, cfg=cfg, latent_image=latent_image, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, denoise_mask=noise_mask, sigmas=sigmas, callback=callback, disable_pbar=disable_pbar, seed=seed) File "P:\ComfyUI\ComfyUI\comfy\samplers.py", line 742, in sample samples = getattr(k_diffusion_sampling, "sample_{}".format(self.sampler))(self.model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar) File "P:\ComfyUI\python_embeded\lib\site-packages\torch\utils\_contextlib.py", line 115, in decorate_context return func(*args, **kwargs) File "P:\ComfyUI\ComfyUI\comfy\k_diffusion\sampling.py", line 137, in sample_euler denoised = model(x, sigma_hat s_in, *extra_args) File "P:\ComfyUI\python_embeded\lib\site-packages\torch\nn\modules\module.py", line 1501, in callimpl return forward_call(*args, **kwargs) File "P:\ComfyUI\ComfyUI\comfy\samplers.py", line 323, in forward out = self.inner_model(x, sigma, cond=cond, uncond=uncond, cond_scale=cond_scale, cond_concat=cond_concat, model_options=model_options, seed=seed) File "P:\ComfyUI\python_embeded\lib\site-packages\torch\nn\modules\module.py", line 1501, in callimpl return forward_call(*args, **kwargs) File "P:\ComfyUI\ComfyUI\comfy\k_diffusion\external.py", line 125, in forward eps = self.get_eps(input c_in, self.sigma_to_t(sigma), *kwargs) File "P:\ComfyUI\ComfyUI\comfy\k_diffusion\external.py", line 151, in get_eps return self.inner_model.apply_model(*args, kwargs) File "P:\ComfyUI\ComfyUI\comfy\samplers.py", line 311, in apply_model out = sampling_function(self.inner_model.apply_model, x, timestep, uncond, cond, cond_scale, cond_concat, model_options=model_options, seed=seed) File "P:\ComfyUI\ComfyUI\comfy\samplers.py", line 289, in sampling_function cond, uncond = calc_cond_uncond_batch(model_function, cond, uncond, x, timestep, max_total_area, cond_concat, model_options) File "P:\ComfyUI\ComfyUI\comfy\samplers.py", line 265, in calc_cond_uncond_batch output = model_function(input_x, timestep_, c).chunk(batch_chunks) File "P:\ComfyUI\ComfyUI\comfy\model_base.py", line 63, in apply_model return self.diffusion_model(xc, t, context=context, y=c_adm, control=control, transformer_options=transformer_options).float() File "P:\ComfyUI\python_embeded\lib\site-packages\torch\nn\modules\module.py", line 1501, in callimpl return forward_call(*args, **kwargs) File "P:\ComfyUI\ComfyUI\comfy\ldm\modules\diffusionmodules\openaimodel.py", line 627, in forward h = forward_timestep_embed(module, h, emb, context, transformer_options) File "P:\ComfyUI\ComfyUI\comfy\ldm\modules\diffusionmodules\openaimodel.py", line 56, in forward_timestep_embed x = layer(x, context, transformer_options) File "P:\ComfyUI\python_embeded\lib\site-packages\torch\nn\modules\module.py", line 1501, in callimpl return forward_call(*args, **kwargs) File "P:\ComfyUI\ComfyUI\comfy\ldm\modules\attention.py", line 693, in forward x = block(x, context=context[i], transformer_options=transformer_options) File "P:\ComfyUI\python_embeded\lib\site-packages\torch\nn\modules\module.py", line 1501, in callimpl return forward_call(*args, **kwargs) File "P:\ComfyUI\ComfyUI\comfy\ldm\modules\attention.py", line 525, in forward return checkpoint(self._forward, (x, context, transformer_options), self.parameters(), self.checkpoint) File "P:\ComfyUI\ComfyUI\comfy\ldm\modules\diffusionmodules\util.py", line 123, in checkpoint return func(*inputs) File "P:\ComfyUI\ComfyUI\comfy\ldm\modules\attention.py", line 625, in forward n = self.attn2(n, context=contextattn2, value=value_attn2) File "P:\ComfyUI\python_embeded\lib\site-packages\torch\nn\modules\module.py", line 1501, in callimpl return forward_call(*args, **kwargs) File "P:\ComfyUI\ComfyUI\comfy\ldm\modules\attention.py", line 420, in forward k = self.to_k(context) File "P:\ComfyUI\python_embeded\lib\site-packages\torch\nn\modules\module.py", line 1501, in callimpl return forward_call(*args, **kwargs) File "P:\ComfyUI\ComfyUI\comfy\ops.py", line 18, in forward return torch.nn.functional.linear(input, self.weight, self.bias)

@epinikion Thanks, I didn't realize.

Doesn't that tell you something? I rarely do girls, but I get awesome results for some things at 17 steps. I find the more steps, the more detail I lose. A detailed, interesting roundish planetoid, with enough steps, becomes a polished steel ball. And I rarely use just SDE. I use DPM++ 3M SDE Karras or Restart mostly. Sometimes exponential for an in-between look. I used DDIM for the longest time because it generates the most details and is most creative. Too much so, in a lot of cases. DPM++ 3M SDE can work for some of my more chaotic LoRA, but it's also the first sampler to turn something else into a stock girl face.

@epinikion Thank you :)

@epinikion thanks,but i figure out problem is with Adetailer =_=,

also some loras or embeddings tend to reproduce the same faces. try to reduce their strength, or give some additional details. instead of "a woman standing on street" add details like "a woman standing on street, {job} {nationality} {name1 name2 name3} by {name of artist}" or add facial details "a woman standing on street, snub nose, etc.", this could give different results. you can also have a look on dynamic prompting with wildcards, this automatizes the process of adding random stuff to your prompt. if using comfyui you can do even more advanced stuff by using different sampling steps and inject detailed prompts inbetween

@OneBullet thank u for suggestions! i'll try them

@OneBullet Any particular wildcard loras/ to try?

@Omikonz  in a1111 i have "dynamic prompts" installed and here on civitai you can search for "wildcards", i have "billions of characters" downloaded (yaml files that go into the extensions folder). and for loras, i havent found a good one that gives a good variety of different faces. imho the best and easiest way to achieve a variety in faces is, to use an ipadapter model. i dont know how to use it in a1111, but in comfyui its really easy. between the checkpoint loader and the sampler you can insert an ipadapter node where you can load a face. this "image prompt" gets then mixed with your text prompt and gives you an image, where the face has similar features to the ip image. you can also batch multiple images together and this "mix of faces" gives you an unique new face, or you can play with the strength of the adapter to adjust the influence of the image prompt. there are also other ways to change the image, i.e. use two advanced samplers in comfyui, the first renders the base image in i.e. 8 steps, and then you inject a new prompt that specifies i.e. the face and the second sampler continues with the injected prompt from step 8 on. then you can change the injected prompt to generate different faces, (add ethnicity, describe facial forms, add names etc) and keep the basic composition of the image. there are maybe other methods, but this are the easiest ones for me.

you can probably get around by specifying the nationality of your character.

@epinikion Ok, damn I did not expect it to be as simple. Still wondering how it actually works, but thanks man gonna try this out.

@epinikion The inpainting models that you have uploaded are merged with SD 1.5 models?

@azirizvi999375 Its not just a normal merge, its adds the differences to the base Inpaint model. I assume adding the weights to this model it translates also to the depth maps, that is the only explanation i can come up with and it kinda makes sense.

And it seems all the custom inpaint models are made like this, because in months I could not find a detailed explanation how this could be done by training. The only reliable Info I found was a paper that explained the model contains depth maps, basically comparable to Control Net depth model, which does work in a similar way.

Wow these model-mixers just can't win. Low-effort prompters complain about the same type of face (while actual young girls IRL are ignored because they don't have that same type of face). And other people want to reproduce example images 1:1. I really don't get this, other than in the few cases where someone is posting images with exceptional detail and content and some people are calling foul. I duplicated one, once, to settle a bet. But you do know how easy it is to miss an embedding? You do know sometimes you can't just read the image gen params in Civitai with your eyes? You have to do the copy button and paste it somewhere. Then you often see way more stuff. Also, I know guys who just got a new graphics card and can't duplicate their own images. Also, use xformers? Or anything like that? Most optimizing algorithms are somewhat 'non-deterministic'. Can we make some kind of sticky for this?

@parallelepipedon, you could've written all of this without being so pretentious, condescending and coming off as high and mighty, I'm very well aware of the elements that change an image and I had trouble replicating his, hence why I asked the creator if he was launching his UI with different web arguments or he was using a specific ENSD, even having arguments when launching the web UI such as medvram influence the result greatly, your comment is unwarranted but you if wanna comment more to satisfy your ego you go right ahead.

@parallelepipedon from an engineer/scientist standpoint this question makes sense. Stable Diffusion absolutely can be deterministic, and if a user has come to this site to explore the computer science behind AI, this is not only a valid approach, but a prudent one.
I can understand aversion to the question from an artistic standpoint, but I think the inquiry was issued by someone looking to validate their installation and results.

@parallelepipedon awful comment

This is better at easy prompt

I can't get any of my loras that work with other models either. Really wish I knew the issue.

RC is "release candidate", so sounds like a sort of work-in-progress but good enough to get out until the next version

@jspapp Ooooooh... Duh...thanks, lol.

Try a different sampler

@epinikion @mailzadev  Thank you this works, and beautiful faces now. I'm still pretty new to this and this is the first model for which I have had to change the sampler. I'm happy to do so because the result is very nice but is there a reason why the default doesn't work for me?

@thakakarot Please share which combination of sampler, steps, CFG and scheduler you used. I am facing the same issues.

I had the same issue when trying to remix image 2384151 (NSFW) and making any subtle changes and also adding the Add Detail lora back in. (If you leave things the same, the image generator just shows the cached images.) The images would have artifacts that looked like bright rainbow metal on parts of the body.

I made the following changes:

- increase steps from 20 to 40, which removed most artifacts but left some weird color artifact in a small part of clothes

- change the sampler from DPM++ 2M Karras (what I typically use for most generations) to Euler a, which removed remaining artifact

I had also tested reducing CFG from 6.5 to 5.0 as recommended in the model . That alone removed artifacts, but not as much as increasing the steps.

After doing some more testing, I found that increasing steps to 50 makes artifacts go away. I can keep the sampling model and CFG of the original image. The model page does recommend CFG 5, so it may be good to use that value anyway.

You can't. Fooocus only supports SDXL models, this model is based on SD 1.5.

nipples everywhere🤣

(((NSFW)))

Try different prompts with races like asian, caucasian or white and mix famous people's names even if they are male to try to get a different face, also use adetailer

Go into advanced settings, look at them :D then find the one that states something like "dev mode, force model swap" set that to 1 so at 1 sample it'll swap models to whatever you have as the refiner

lora, ipadapter

Don't use textual inversion embeddings, such as EasyNegative. While helpful, these are the #1 cause of sameface problem. If you still don't get enough variety, try adding "Nigerian/Irish/Portugese/other country" to the prompt to get a specific type of face, changing the country every few gens.

The only known way to stop that is to merge the block weights of a model that produces different girls, like DreamlikePhotoReal or LiberteRedmond, though that comes with problems of their own, or you'd be using those models instead.

I use Reactor to produce multiple same faces looking to different angels and then trainer a Lora with the model.

use a lora that makes different people I can't remember the name but its right here on civit

@DobleMM RealHumans, you mean?

I was dealing with this all day today even when using embedding:epiCPhoto-neg the negative prompt I ended up using that worked most effective is (nude:1.5), (NSFW:1.5), naked, breasts out, nipples thats the only negative prompts I use with this model and I get good results now

I've been using some other models @astr010. I guess epic is geared towards nudity. Really liking Ican'tbelieve_newyear... I've also started using LCM models and have been getting great results.

@bigmax8999 Bro thank you so much for the tip I am taking a look at these now!

@epinikion ok will do. I was experimenting with the same prompt across multiple models and epiCRealism has consistently left the cloths out even when explicitly detailed in the prompt. I will test with epiCPhotoGasm as well

isnt for me eithe

Tech people just love it when people complain something's not working. Is the developer supposed to guess how it's not working? It could be anything from actually throwing up an error because the file got corrupted during download to it doesn't work with your prompt the way you think it should.

Did you try unplugging your computer and plugging it back in and see if that fixed it?

it's that way for all the checkpoints. to change things up, put the name of some celebrity, but inside square brackets so it's not too much. it should radically change the face type, if it is a face from the old days, put 'monochrome' and 'black and white' in the negative prompts. that ought to get you started.

@fablegenius Thanks for the tip!

The best solution is to use wildcards. Firstly, use chat gpt or something similar to generate names of different countries and ethnicities, and then use those as wildcards to randomize your future SD generations.

example prompt: a beautiful 34 year old __ethnicity__ woman

@ImAbbieKitten I've been using European, Swedish, Danish, etc, but even then I think it will use the same face for each country/region. It's a shame there isn't a keyword in SD that randomizes the face specifically. We all know for a fact a lot of these models have learned hundreds of different faces, it's just very strange it prefers to only use 1 while everything else gets randomized. I guess it's just another fluke like the bad hand/feet anatomy.

@EROS1CA Another method you can use is to use different names such as Emma Muller, Amber Grey, Priyanka Singh, Xia Lee.

Names like these will surely impact the generation. Try it...

@EROS1CA  I don't know why but it's true that most faces tend to be the same, almost regardless of which checkpoint. but if it's face types you want, i stand by my recommendation over other alternatives such as race, country, etc. it can be a long process to find what you're looking for that way, but you can try some tests with the s/r function in the XYZ grid if you're using a1111 for example. load up a bunch of celeb names and tame them with the square brackets so the standard faces get blended with the specific faces. but sometimes SD might not "know" the celeb you're naming so it might not react in those cases. trial and error.

@fablegenius I did unknowingly combine 2 celebrities before I started noticing this issue. You're right it seems to do a great job at creating unique faces.

Check your negative prompt for negative embeddings. Negative embeddings (i.e. EasyNegative) is a fast way to start using checkpoints, but they are likely introducing their own biases to various features like faces.

@EROS1CA that's good to hear. it seems to me that adding a little essence of various celebrities is easier than messing around with LORAs in most cases. Don't forget about Adetailer as well, it's a must-use feature.

Using a celebrity's name works the best for me to get a different face. You could do different names too which should change it as well.

no, these types of models have not been trained for this, you need to use LORA

@epicPhoto it sure knows how to make a vagina tho.

@epinikion and for the lazy folks? :))))))

with those couple sentences, your English is clearly and sadly better than most native English speakers 👍

@epinikion Unfortunately, the noses are getting wider, not longer. But if it would be useful to anyone, I found that TDLX_Nosesize LORA helps to get control over this parameter

yes - it's just a checkpoint - it will work with any SD1.5 workflow you already have working in comfy.

https://civitai.com/models/277058/epicrealism-xl