@coyotte Thanks for the help/feedback. I love your models, please keep going.

correct.. foocus just patches a normal sdxl model so it can inpaint, but you can also use standard illustrious models too

@coyotte ok, thanks! I was using denoising at 0.5 and 'latent nothing' for the masked content (for big changes in the image), and it wasn't too bad, even with the dmd2 lora (LCM Exponential - cfg 1 - 8 steps)

@coyotte Thanks for the reply. In that case, I'll do the merge and test it out myself!

@coyotte thanks

I kind of liked the smoother texture version. Is it still possible to download somewhere?

@kunde2 sorry i missed your comment. yes, he still has it up there, it's just moved to the absolute far right of the list.

@coyotte if you look at your comment section, you will find a couple more "for me" cases. No other inpainting model has those comments. "bad results" means you get "noisy" inpainted areas.

@coyotte The results this checkpoint produces are worse than any 1.5 ( i know that this is xl) inpainting models I have used in terms of detailsa nad every other aspect. 1.5 models are not only better but also faster. Not to mention the noise in the inpainted area. Unusable.

@coyotte Obviously not enough about inpaint models

@coyotte Are you serious? Even on that screen it shows how it produces low-quality results. Even on that featureless skin. The latest cyberrealistic inpaint (mind you 1.5 model) is light years ahead of your model. Actually, all of the earlier versions are also. There are too many one-click inpaint models around here, spamming the feed for no reason. Your model got deleted. The only reason I am still writing are your comments spaming my notifications.

@ledeni80 I've been to circus so many times, yet this is the first time I see this type of clownery.
I had to download the checkpoint just to see what this all was about. Not only it worked without any tweaks, but it also outperformed my go-to checkpoint (Juggernaut Inpainting). So... thanks for bringing my attention to this one, I guess? @coyotte Seems like this dude's just a pathetic troll trying to feed on your reactions. Better to ignore them completely

@coyotte I haven't inpainted in awhile. You guys really using comfy directly? Isn't that a pain to mask in another program and then manually load image?

@ledeni80 Tried his workflow after a mistake in my end and it works fine with decent quality if your prompt is fine.

@Lena_Luna Well if your "go-to" is Juggernaut that explains a lot actually.

@ledeni80 ?

896x1152 

832x1216

1024x1024

1024x1536

I have been getting better results with some practice and step 2 seems useful I will try it :) thanks

Not sure about this model, most models have vea “build in”, if image looks really bad then probably you need to use separate vea. Just try one with and one without. Hope that will help you 😁

There are no optimal negatives. Personally I don't use any. Negatives were very important with SD1.5 for improving quality. Not so with SDXL.

@SubtleShader thanks!. I'll continue using no negs with your model.
I do think though they might have an affect on certain models, atleast i have come across illustrations as output instead of photorealistic unless negg'd. also some other wonky renders without negs in some other models.

@coyotte Congrats on your first lora!

Try this prompt after your positive prompt:

, ((photorealistic)), ((RAW photo)), ((8K resolution)), ultra-detailed, ultra-sharp focus, realistic skin texture, visible pores, subsurface scattering, lifelike skin tones, natural facial expressions, realistic imperfections (fine lines), accurate color grading, true-to-life tones, natural composition, authentic details, Nikon D850 DSLR photo, full-frame sensor, masterpiece, award-winning photography, best quality, (realistic eye highlights)

and this one as your negative prompt:

(worst quality, low quality, bad quality, poor quality, terrible quality), (bad anatomy, wrong anatomy, broken anatomy, malformed, mutated, disfigured), (distorted proportions, awkward pose, uncomfortable pose), (overexposed, underexposed, oversaturated, over sharpened, overfiltered, bad lighting), (lowres, low dpi, pixelated, compression artifacts, jpeg artifacts, aliasing), (blurry, motion blur, smudged, rough sketch, unfinished, messy), (extra fingers, fused fingers, missing fingers, deformed hands, mutated fingers, twisted hands, missing hands), (extra limbs, fused limbs, disconnected limbs, missing limbs, abnormal limbs), (ugly face, uncanny face, worst face, deformed eyes, asymmetrical eyes, (cross-eyed:1.5)), (plastic skin, waxy skin, doll-like, CGI, fake skin), (tan lines:2), monochrome, grayscale, cartoon, anime, sketch, illustration, vector art, emoji, meme, watermark, signature, artist name, placeholder, text overlay, logo, sample watermark, instagram filter, censor, censored, mosaic censoring, obstructed view, style mismatch

@coyotte Well try it, and judge later

@vicaut You are funny. I'll give you that. 🤣

@coyotte and i use these prompts not only for your checkpoint, but for all SDXL and SD 1.5 checkpoints if i want more realistic pictures. There are even more prompts i use like cinematic lighting, soft ambient lighting, shallow depth of field, volumetric light, realistic shadows and reflections, ray-traced lighting if there should be a light soure involved - also works great. I even use special prompts for hires fix and adetailer. If some prompts are useless for some checkpoints, no problem, at least they won't do any harm to the quality.

@SubtleShader well try it, and judge later. This will work at least for better realism (not for an amateur picture look).

@makaveli_313 just try a prompt with the same seed with and without my addidional prompts and compare.

@vicaut Sorry, but I won't use that mess of a prompt. Hard pass. Robs my prompts of tokens that I can use better for other things. I admire your confidence to suggest something to people who actually tried thousands of prompts when creating their models. Including the stuff that you are suggesting.

@coyotte oooh... I will try this later today. Would it work fine on olt fixed textures model too?

@coyotte thanks ill provide feedback later directly on lora page.

@coyotte No hurry and no obligation to contact me here or there. Just thought we could discuss "joining forces" somehow.

You don't need a special DMD2 version. Just download the DMD2 lora from https://huggingface.co/tianweiy/DMD2/blob/main/dmd2_sdxl_4step_lora_fp16.safetensors and use it together with the chrckpoint.

Yes, much more pleasant to work with the DMD2. Much faster, less anatomy problems and usually looks more photorealistic. Without DMD2 use the SDE sampler with Karras scheduler.

@SubtleShader What are the best settings for this in Ur opinion

@coyotte do you have any idea what could it be?

@coyotte Thank you for the quick reply. Indeed, after your comment, the fix was easy: just load the model as follows: from diffusers import StableDiffusionXLPipeline

pipe = StableDiffusionXLPipeline.from_single_file("lustifySDXLNSFW_oltFIXEDTEXTURES.safetensors")

@lucasnuzdas4481 This is really great code snippet ,thank you! Could you please provide further references to develop further scripts? Thanks

try prompts such as cleavage or make sure the girl put some clothes on

try prompts such as cleavage and buy her some clothes

@coyotte Thank you for your reply.

@coyotte I agree it is inherently different, so maybe releasing it as something else would have been the right move here, that is simply my poorly informed opinion. You released something inherently different under the umbrella of something unlike it. To me, it seems dishonest to your artistic prowess, and the possible legacy of the definitive version of Lustify which is Endgame and not OLT.

TL;DR You are the artist, don't obfuscate your creations with one another is all I really am trying to say. Both amazing executions

@coyotte would you consider going 'in the direction' of Endgame for your next one?
Thanks

@Yushio What do U think is not in OLT that is in Endgame that U think makes OLT different and Endgame much better ?? I am using both but I am just more interested in what Ur opinion is further in elaborate detail
Thanks

@sam48052023878 OLT doesn't have something missing. They are inherently very different. If you look through the OLT and the ENDGAME Generations you can visually see there is a clear distinction between the two. OLT images look like they were taken by a professional photographer and should be on Instagram or some gallery no matter the quality of the image. While not being as stable. Endgame is rough, It could be used with ease with SFW with little tweaking and it was very stable. Images can be generated and look like a total noob at photography or someone who just picked up a phone camera.

OLT has a more refined commercial feel.
Endgame feels more malleable and rough.

There aren't missing things, they are just different.

@Yushio Precisely described. I can't speak for most of the people but I strongly feel Engame was the correct direction.

@coyotte Ah ok, I just assumed the performance of the denoising strength varied greatly between different upscalers. Thank you

what are you using comfyui? or on civitai i use comfyui and with the right settings as mentioned in the info its good

IOPain + choose lcm lora (built in) and you fixed!

@png i'm having the same problems as OP, what is "IOPain" where do i find the lcm lora?

Nevermind I'm a dumbass. I thought Denoising strength is meant for literal denoising (like a post processing step) and set it really low.

i'm new too. so do you put denoise all the way to max level then? do you change your prompt or keep it blank?

@coyotte can you help me with this question i want to know if V6 is more realistic compared to v5 endgame

@sonic4life170  I would say in terms of realism, for me at least, its hard to tell much of a difference, both are so fkn good!

@ronikush yes thats true there is not so much quality difference i found out that the textures are better in the V6 model, but coyotte said something about the quality but still its a bit confusing cause there is a difference between realism and realistic....... i agree they are great its the best model i am still using endgame. cause i am not sure about the super realistic outcome

For me I have most success when I cut out using negatives completely and if going for SFW try to describe the clothing in as much detail as possible

whats your recommended settings/workflow to get good img2img results?

whats your recommended settings/workflow to get good img2img results?

@atomo111 
Hi, a typical workflow when I use img2img can look like this (Keep in mind though that my workflow is set up with the desire to create a realistic image. It could be completely different if you want an artsy image or painting style):

1. Uploaded image size: Mostly equals the default size in the generator (1024x1024 etc). I often work with a cartoonish/anime stylish picture where I've got the basic desired pose, angle or thematic element (I use NovelAI to create this "base image" because I feel comfortable with its quick rendering and prompting language and especially the ability to use an Inpaint tool, but in theory it should work with using any Pony/Illustrious checkpoint to render a base image on this site as well, though I can't recall myself having done that.)
2. CFG scale: Almost never higher than 4.0 - it's easier for a checkpoint for realistic images to render a more unique image with a lower value from my experience.
3. Sampler: DPM++ 2M Karras if I wants more attention to details, which I usually do (But always read what the creator of the checkpoint has suggested) :)
4. Steps: Usually 30-50, I like to experiment with this. Lower can work too (especially if I switch sampler to Euler).
5. Denoise: About 0.45-0.60. This depends on the base image, sometimes, if not mostly, a higher value increase aesthetic value (from my personal view), but can lead to undesired content such as bad hands, eyes etc. I often do at least two images with different denoise settings to see what effect it has. Loras can also tend to add undesired content, and I try to change their strength to see what that leads to. I find myself often use a slightly lower denoise level when rendering the first image, and then I increase denoise when doing the hi-res fix.

The generated image I find to be good enough is often put through hi-res fix unless the first one already is great. Step 2-5 above are repeated, with eventual adjusting in step 4-5. NB: The hi-res image sometimes fails! Bad eyes and bad hands can be a result (hands are often quite challenging). I try to adjust denoise and steps and retry. Sometimes I say "Fine, the 1st one was good enough, post and move on," to myself. It can even be caused by the base image simply not being suitable, and I have to discharge it completely after several tries.

Lastly, I use face fix if needed. It doesn't always work, it can worsen the quality, but it can fix blurry or deformed eyes or strange skin hue. I experiment with different kind of skin detail loras to see if they help with this.

About Loras: If you are uncertain about loras - Always begin with only using the checkpoint.
Then add loras one by one to see what they add/change until you find it to be satisfying (this is perhaps the most fun part - to search for those hidden gems around on this site!). Don't throw in five, six loras at once, it makes it difficult to exclude which one causes undesired content. Do also take a look at the checkpoint site if the creator has added suggested resources, helpful in enhancing image quality, or check the published images by other users to see what loras they've used and with what strength, depending on what type of image you are trying to create.

All steps above are done usually with the desired prompts, checkpoint and loras from beginning once I've gradually built up the resources I like. I can switch loras before hi-res, just to experiment.
As you probably realize, this type of workflow demands you commit some time o it and it consumes a lot of buzz, and I fail (a lot). It's not a perfected workflow by any means. But a famous chef ones said something like "If you fear to fail you should never try to cook food".

I hope this was helpful as I tried to be as specific as possible. It turned into a whole novel, sorry! I am far, far from an expert.

@Tapiola85 This is so incredibly helpful thank you so much!! I've been really struggling with img2img so your response is invaluable to me. So kind of you to lay it all out. Thank you for also mentioning the lora's, most of the one's i've been using haven't been working well on this checkpoint but lustify is my fav checkpoint ever so i'm going to play around with them a bit more to see what works best. I'm looking for super realistic images with film like quality. Thank you again. I'm saving your reply

@atomo111 Glad I could help! :)

yes both these models have difficulty at creating anal penetration but on the other hand DMD2 makes anal images as it is no business, try DMD2 and lemme know does that work for U

@sam48052023878 Thanks for the recommendation but it didn't work for me. I used all sorts of prompts and weighting with both Booru tags and natural language, but it didn't produce a single one.

What prompts and workflow are you having success with?

how to use this with fooocus can u teach me?

@kr96aditya I have no experience with fooocus sorry. I use ComfyUI.

interesting, what do you mean by highres fix? it's a lora ? ... what is your ksampler config?

@mfudi I'm using ComfyUI. Highres Fix is a node from here: https://github.com/jags111/efficiency-nodes-comfyui