Works for me

Looks like a penis is coming out of a man's butt in one of the previews.

this is a LOKR/LyCORIS, not a lora, thats why its breaking for me. Please classify it as the right type!

@Ashen3 It's actually an optical illusion false alarm. Once I zoomed in I saw that it's just the light on his leg. It's the one of the guy holding the woman up in the air.

check again bro

@ElGordoAI check where?

If you mean the file posted to this page, it's a Lora. I meant a model finetune like they have for Klein.

bro do you have basic workflow? have the same problem and dont get it how to solve it

@Seki707 just copy the OP's workflow (copy nodes when viewing one of the images and ctrl+V when focused on comfyui). I changed the steps in each sampler a few (the first to 36/0/6, the second to 20/6/999), and decreased the turbo lora on the second sampler from 1 to 0.8. You'll know the turbo lora is impacting your outputs if the color changes from the first sampler to the next. Your biggest impacts however may be the prompt. Make sure to reduce negatives, and to minimize quality references unless it's specific. The easiest way to remove the art effect is simply "raw amateur candid photograph of".

@makiaeveli thx buddy

@makiaeveli the "raw amateur candid photograph of" saved me thanks I was getting alot of anime even with negatives

@dxjaymz Using anime-based negatives can increase the anime style. Also, using vague words like "cinematic, atmospheric, dramatic" can increase the anime style. Words like "crisp, unedited" seem to catch it too.

I noticed this too, output looks better regardless content, also it corrects prompt adherence in positive way.

Here: https://civitai.com/models/2727641/krea-2-turbo-lora-256dim

Sorry I don't get it. What's/Is there a difference between "krea2_turbo_lora_rank_64_bf16" and "krea2_raw_to_turbo_r256_comfy" loras?

@Ashen3 Thanks for the reply. I was just realizing you use it the r256 for the second/turbo pass and the rank64_bf16 for the normal pass on 1/4 strength. Therefor I thought there might be a different (important) training approach in the two models.

the example workflow is with raw, they just have a turbo lora on the second sampler to reduce the step count. You can turn off the turbo lora and increase the step count to 30-50 and see similar, probably better, results.

generally using the word "nipple" is the issue. finding other words like "pokies" or not referencing "breasts" at all. Generally I use terms like "appropriately covered" if I'm getting really annoyed

@makiaeveli yeah I'm not using any of those words. :(

@pogo Lowering the strength to 0.5 can also help, or at least lowering it to a point where the nipples won't force itself through. It's also worth seeing if the base model itself is placing the nipples, as I'm pretty sure it can. I'd just keep talking about the shirt and the top, it almost always gets me at least some good results. Otherwise the options are using masked inpaints in other models to simply remove the nipple.

Try less explicit words, like 'bust' or 'cleavage'. Also, reinforce that the person is clothed. I often do 'large bust fully covered by her top'. Seems to help a bit.

From another comment, that happened as well as anime styles: I figured this out, it's your noise schedule. Increasing the turbo lora strength was my easiest fix, but if you're using the two-step workflow, make sure the samplers have better "steps, start_at, end_at" settings. I'm using the 36, 0, 6 -> 20, 6, 999. The second sampler has the turbo lora at 0.8 strength. Both using SNOFs at 1 strength. That completely removed my anime styling.

@makiaeveli, are you using base model with turbo lora, or turbo checkpoint with turbo lora?

@vamorand The base, however that didn't reduce the amount of anime I was getting, but it did improve my overall results. And this turbo lora: https://civitai.com/models/2727641/krea-2-turbo-lora-256dim. The actual turbo checkpoint probably doesn't need another turbo lora. The whole point is to just direct the raw model towards outputs that the turbo would produce anyway.

@makiaeveli, I'm using turbo base model already and getting weird results on small prompts. Everything tends to be of some monochrome traditional japanese ink draw style. It can be suppressed if you try hard enough, but that still means that lora messes with the result much more than it should.

I just need a good nsfw unlocker while preserve amazing krea 2 capabilities. I tried other options like 2 comfyui nodes, but they are lobotomize the model too much. Waiting for other options I guess

@vamorand any time i get it now it's just my prompt. adding in some variance of "amateur candid photograph" tends to give better results. if you're using tags instead of sentences, that could bias it towards anime

I figured this out, it's your noise schedule. Decreasing the turbo lora strength was my easiest fix, but if you're using the two-step workflow, make sure the samplers have better "steps, start_at, end_at" settings. I'm using the 36, 0, 6 -> 20, 6, 999. The second sampler has the turbo lora at 0.8 strength. Both using SNOFs at 1 strength. That completely removed my anime styling. (edit: i said increasing the turbo lora, cause I was testing it at 0.5 -- however my original tests using OP's workflow is using it at 1.0).

@makiaeveli are you certain about those settings..? 36 steps total on the first with and then 20 on the second with the turbo just gives me latent noise :D

@bewunderteradler112 the point of the 36 is to make it only 6/36, so it aims to have a noisy image in the first step. The second one should pick up where it left off, aiming to finish in 14 steps. So 20 steps in total. The point of the first stage is to have a noisy unfinished image to pass on to the "real" finisher, to give it a more random starting off point each time.

@makiaeveli do you change samplers or scheduler in between? I want to try to replicate that flow.

@bewunderteradler112 it's just the OP's comfy workflow with slightly changed steps. Really all I'm adding is the extra 2 to 4 steps in each sampler, and reduced the turbo lora from 1.0 to 0.8. Essentially extra steps to mitigate the lowered strength of the turbo lora. The easiest change is simply adding "a raw candid photograph of" to your prompt and removing any other quality references. The negative should also not have any references to drawings or anime, especially since we added the extra steps to the first sampler with the 4.0 cfg. (to clarify though, just these extra steps were enough to bump some of my prompts into realism, but more stubborn ones needed the prompt buff. It's likely because of the turbo lora's effect, so reducing it and "stepping it out" seems to reduce the effect). (to further clarify: you could try using different samplers at each step, that is a common Z-img inference strategy: use the expensive sampler ala res-2s in the first, quick, noisy result, then use the cheap sampler ala euler in the sampler that should resolve eventually).

@makiaeveli really appreciate all the inputs! thanks :)

I have this issue, but to an extreme, using the default ComfyUI workflow. Without the Lora, I get a character that respects my prompt: a photo of a woman in her 30s. With the Lora I get a weird mix of 3D anime/cartoon (a mix of both) render, and she looks extremely young, like "teens" young. This is not the only 18+ Lora that does this, btw, but it is the worst one. It's a great Lora (way better than the filter removal one) when it works, but this quirk is real and weird.

@MrCylion I was getting this with the Turbo base model, especially when I was using short prompts, or tag-based prompts, and oddly even at Portrait resolution but not Photo resolution.

Swapping to the raw base model removed the differences between the resolutions, but slightly increasing my steps in each sampler fixed a few of the longer prompts. Keeping the turbo at 1.0 instead of 0.8 fixed a few others. I also found that referencing "anime, cgi, 3d", etc in my negative prompt also created the drawing style (from Danrisi: "cinematic, glossy finish, shallow depth of field, cinematic bokeh, uncanny anatomy, frame-perfect symmetry, blurred background, fat, big lips, thic lips").

But fully updating prompts of the shorter ones, even with the "photo-based" negatives, fixed it almost entirely. I only get anime stylings rarely now, and fix it on those particular prompts by either turning on photo style lora (like lenovo or garlands or vanguards, at 0.8 strength), or fleshing out a prompt to make it more realistic (at the least: "candid amateur raw photograph of... Natural lighting, candid amateur perspective.").

I think since the model essentially has all the anime support out of the box, you really need to push it with loras and your prompt to keep it from popping up.

Ok sorry for all the spam, I just wanted to make sure it was all in one spot.

@Ashen3 please do it, this will be super useful to use with INT8 ConvRot quantizations which are becoming very popular now that ComfyUI has added native support, but large LoRAs may impact generation time due to the on-the-fly merge.

@Nexdoor it would probably be safer to have it all in one repo, but it's quite likely these scripts can just run inside your comfy directory with no extra install, at least that's how I'd normally cast other models down

merge is not authorised, so I'll see if the author allow it. but you have the python script on my lora page if you want to do it locally :)

@Nexdoor got SNOFS greenlight so it's up ;)

faces feel like an impossible task to keep unaltered unless the model trains none at all -- which would similarly be difficult to convey all the topics. Even using the raw model, we can tell it knows characters for example, but the answer is having a separate small character lora to guide the characters back to where they should be, you could even alter the noise schedule to come in near the end with a much lower strength of snofs + a higher strength of the character lora.

You can use weights to override it if you want a particular face that you can generate without it experiment with adding (central eurorean face:2.0) or however you describe the face normally (description:[number from 0.1 to 2.0]). Also keep the weight down sub .8 (I usually run at 0.7) for SNOFS. But otherwise it's actually really good. Also try adjusting eye positions (it seems to change the entire face shape), eg Eyes half open.

okay, the lora power loeader doesn't throw errors if the loras aren't present ;P

@WhatTheGuy make sure everything is updated, I haven't see any lora load errors. I found steadily increasing the step count of the second sampler until it finds a good outcome, I've increased to 26-36 steps for most of my gens

SNOFs is all that's keeping my Krea workflows afloat rn, so many of my prompts completely fall apart without it. Realism Engine catches some of these extreme edge cases. The F2K versions didn't like cowgirl sex, the Krea version seems to not like blowjobs. Still crushes though lol.

@makiaeveli https://civitai.red/models/2749801/rough-stuff?modelVersionId=3093366 this will fix your blowjob problems use it whit snofs

@Addictedd good call out, pretty much forgot about this one. I tend to use either dicktator or rwflaccid at a low weight, they tends to at least get the anatomy consistent

Krea is like that. You should have KJNodes with upgrade and add 'Krea2 Prompt Weight' then with it add (anime:-1) and finally got all realistic prompts.

don't add "anime" to the negative prompt lol. You're essentially triggering it to think of anime. Instead, make sure your prompts are in natural language, or at the least have a sentence at the beginning with "raw amateur candid photograph of". You could also try "professional photograph taken with [insert real details here]". Use negatives that are more photo-based, like Danrisi's "cinematic, glossy finish, shallow depth of field, cinematic bokeh, uncanny anatomy, frame-perfect symmetry, blurred background, fat, big lips, thic lips,"

If you mention just one "anime" negative, now you have to specify all of them. It's much easier to just say none of them. That goes for the positive as well. Most of my prompts that turn "artistic" have vague descriptions like "atmospheric" or "cinematic", are often just tags or comma-separated phrases instead of sentences. Just clean up the prompts and they work.

I'll try.thank you guys