This

use NSFW in positive prompt...

yes you can

I noticed that too, I'll stay on lastfame for now

Not sure what your settings are but I've always used lcm sampler with a low cfg (between 1 to 2), and when using a really low cfg like 1-2 the negative prompt has little to no effect.

@barry9000 Thanks for the hint. It's the most basic workflow, just to keep things simple. I'll try to use different positive prompts then.

not having any luck in Comfyui - most images are coming out weird. Does anyone have a workflow to share?

@epinikion I noticed when I clicked on the models on TA it says "The model already exists. If you are the creator, please contact us to claim it." Mind you, I've been reporting sexualised images of minors to them for a while now and they keep saying they'll do something about them, yet months later they are still up.

@epinikion do it regardless brother

@epinikion Inpaint details is far from perfect but you can do it with small denoise and multiple steps. Replacing large areas is more or less imposible.

I've been using it in SillyTavern with a pre-fixed prompt that includes a relatively detailed character description, and find that sometimes it does very well, and other times the character look changes a fair amount. I have not tried using a LoRA.

Just use the defalt image generation and don't put anything in negative. SDXL needs almost no prompts. Euler a normal - 5cfg, 30steps. Impossible for that to produce a bad image.

i use a slightly modified version of this workflow:
https://civitai.com/models/1663229/lwf

cfg usually around 4.5 but i have gone up to 7 when i need to force some prompts. been getting fairly good results, it doesn't need neg prompts but even if you give them it shouldn't have to much of negative impact, for the sampler DPM++ 2M (Karras) is my goto.

sometimes it just works pretty well even nsfw, its just combination of prompts thus i believe and one must try to experiment i must find out more precisely but there might be many "random" rules, after all its ai, thus its kinda picky, it thus might be not some programmer hardcoded rules but ai deducted

On crystalclear, add NSFW in the prompt. LastFame dont need, rather, it should be slowed down with prompt negatives...

epinikion  lastfame dmd2 literally every single one of my prompts has clothing details i tried to generate images with various outfits even turtleneck sweater prompts give nsfw results like nips out or spilling out, i even used clothing loras but it just doesnt work i cant manage to keep the breasts inside lol

fantastic8 fantastic8 doesnt work added nsfw, exposed breasts everything in the negative prompt just doesnt work

dhirenmainali705 And that is bad, how?

not if you describe what type of clothes you want in your prompt. or just say dressed in this color or that color outfit

@jch66 Im doing that and it still wants exposed breasts. Putting different version of covered breasts, and different kind of clothing, "exposed breasts" in negative, heavy weights, tried it all none of it works.

i put naked on negative, work some times

Try PerfectionAmateurXLIL

eoa42 I will try it thank you very much for the tip, I am however more likely to use LCM models because I still use an old VGA 1080ti and the generation is very slow with the script I made...

Shinny oiled skin

This applies to me. I prefer Fooocus for image generation. For a long time I favored IrisLux and then switched to Donutsdeliverymixv3, and these are still among my favorites, but EpiCRealism (CrystalClear) is the best so far. Two setbacks (from my point of view): The eyes sometimes lack integrity, unless you do super man steps. I prefer 30 steps because I'm impatient, but with a detailer and enhancement Styles; the other thing is NSFW, where it tends towards fellatio when I prompt for anilingus. :D

epinikion i am asking about schedular.

epinikion thank you so much

Do you use comfyui? If yes, what workflow?

yes, dmd version

Use the dmd lora.

Which workflow do you use?

mauriziomkr Hi hi Maurizio, if by workflow u mean how I generate, I use only Civi's own generator coz my sweet lappie duzn have d power 2 run SD sadly. My NVIDIA is 1650ti 4Gb. But if u mean wat chkpoints, loras etc I use, chk my pix, all d info is there. Bless ya!

epinikion D THNX goes to u, bless ya!!

SinStim12 Yeah I meant how you generate. Makes sense that you use civitai is your gpu isnt' powerful enough.

SinStim12 I do not know the characteristics of the NVIDIA 1650 TI 4 GBS, but I have the 1060 basic with 3 GBS and I can perfectly generate both in SDXL and in Flux in Forge. Obviously in Flux it takes approx 4 minutes every image, it is not practical, but the fact is that I can.

In SDXL I use the LORA of DMD2, 8 steps, CFG 1, Sampler LCM Exponential, size 1216x832 and generates 4 images in less than 4 minutes. No credit limits, uncensored, without control of any kind.

Your GPU is newer and I suppose superior to mine. Try it with that configuration that I use, I hope it will work for you.

Do you use comfyui? If yes, what workflow?

I admire your honesty lol!
Use LastFame for optimal porn : P

@the__Realist Nice :P

@epinikion okay have fun... hope you'll come back some day and give us another awesome model : )

you can use any model as face detailer in comfyui if you set the nodes right

@awesomesnipee You can use it, yes, but the inpaint models are specialized on getting rid of hard edges between the inpaint area and its surrounding.

text? can it do text?

Training a lora is fast and simple

Step one is realizing that consistent faces or “characters” is essentially impossible with any image model without a lot of time and effort and knowing the TOOLS available to you. You can’t replicate a face without some extra tools like loras, controlnet, instantID, textual inversions (embeddings), etc. ControlNet Union SDXL is one of the easiest ways to get started. Text embeddings are fairly easy to create. Even loras have become much easier to make and can truly lock in a consistent “character”. Unfortunately, just learning how to set up a simple text to image chain is actually the easiest part. All the interesting stuff often requires extra learning and work, but that’s part of the fun.

1. creating a lora (you can do that online on tensor art for example and quite fast and cheap) 2. some models (not tested with epicrealism) creating more consistant faces if you give your model an acutal name like: Jane Doe. "realistic photography of Jane Doe..."

Friends, if you have carefully read the author's introduction, the recommended magnification limit for all XL models is 1.5 times. If you need a higher resolution, you can use the 4xNomos8kSCHAT-L model to magnify it 1.5 times and then use the 2-8x model to magnify the external space.

@hanekawa0917792 I have used 1.5x, 2x, 2.5x, no difference. I have used this model in the past and other models in the series with no issues. So I feel like I am forgetting something that I used to do.

@hanekawa0917792 I have used 1.5x, 2x, 2.5x, no difference. I have used this model in the past and other models in the series with no issues. So I feel like I am forgetting something that I used to do.

@epinikion Maybe you have misunderstood what I said. The author I mentioned is not you. I have been using models from your series and the results are great. Of course, I also use models from other authors. Most authors will explain the parameter settings for the upgrade.

@TomatoAlien Maybe there is a problem with your settings. Do you have an email address? You can send me a private message and give me a workflow to try to solve the problem.

@epinikion @epinikion @TomatoAlien's response was that the model was blurry after enlarging. I asked him to try it at 1.5, because many authors of XL models recommend that the magnification is best at 1.2-1.5. I never said that your model can be enlarged to 1.5 at most. I agree with your later point of view. Maybe he used an inappropriate enlarged model or unsuitable scale parameter settings, and omitted the upgrader model. How to make the workflow run?

Easy solution: Use an AI upscaler, just google it there are a lot, I use two services and I got pretty 4x upscales with no distortions, no quality loss and zero artifacts

@epinikion thank you! Im gonna try it!

RTFM

yes

wait what, I'm a little lost on what you mean by that, but super intrigued. What do you mean exactly?

@bigbearpun11 that he is doing a lora or the Sims 4...

@bigbearpun11 The model prompts perfectly to "Character in the Sims 4", or Skyrim, high octane render etc. using tag Character Sheeting. This has alot of the base of original SDXL model so most Lora can work with cfg stretched to 16 then moving the negative prompts with a low token (low detail, blurry, jpeg artifacts etc). Then move to photoshop/blender etc stretch, very fun to do. Doesn't look like a sprite from the early 2000's pasted onto a model because it blends in the original Sims 4 Textures.