This model have been created, by merging process, to produce NSFW hyperrealism images. It can also be used to create SFW, or at the opposite, porn images.
This model have been created for my proper needs, as I was not really satisfied with actual realistic pony SDXL checkpoints. And I share this model with you.
Recommended parameters :
Steps : 25
Sampler : DPM++2M Karras or DPM++3M Exponential
CFG Scale : 5.0
Clip : -2
Pre-Prompt : score_9, score_8_up, score_7_up, source_photo
Negative prompt : score_6,score_5,score_4, score_1, score_2, score_3, source_anime, source_cartoon, rating_explicit, greyscale, text, watermark, lowres, jpeg-artifacts, strabism, heterochromia, fused elements, low quality, poor quality, bad eyes, bad anatomy, unnatural teeth, normal quality, censored, poorly drawn eyes, poorly rendered eyes, 3D
Our Team is actually bidding 20.000 buzz per week on Indecent. We hope it will be enough to keep it available for futures images generations for you all.
Description
This new version of Indecent, after a pre-selection of 21 checkpoints, is based on the optimized merging of the best 8 tested and selected hyperrealistic Pony checkpoints.
FAQ
Comments (85)
I tried using the settings you have recommended in the description, most of the time I am getting loads of mutations & abstract artifacts on SD Forge. Sometimes I get clean results.
Does the Denoising have to be up to 0.70?
Does your artifact happen before or after highres fix?
@GTSuya_Studio Before & after Hi-res fix on my RTX 4060 Ti, I've also been testing with a wide variety of upscalers.
I'm not sure if my total VRAM is effecting generation quality. My current PC doesn't have as much RAM as my previous build. I'm also using SD Web-UI FORGE too.
@GigaByteWide I only encounter artifacts when I'm using the CD Tuner option on Automatic 1111. It could be also related to the VAE.
Quite good for diverse and good looking characters. Takes care of some of that uncanny plastic pony feel but still way behind base sdxl models. Struggles with futa.
Do you have some example of those (NSFW) base sdxl models?
If I was smarter I would have clarified I only meant behind on the realism. In this sense, the standard afaik is still biglust.
@neuronsactivated Thanks
Tried over 50 different sdxl pony illu. variants, yours delivers excellent, realistic results. Thank you for your work.
Thank you !!!! :)
Very good merge. Who needs Flux??
Thank you
This is a GREAT checkpoint. You know you can bid to get it boosted for a day?
If you don't win, you get your buzz back
Required buzz are relatively high.
I was able to boost it for today with 20k buzz, but I was not able to see what was actually the effect.
Strong checkpoint. Nice examples. Looking forward to testing this. Real skin texture, strong colors and contrast.
Was missing this type in the options until now. Bit expensive now but should rise up in no time.
Thank you. But takes into account that currently civitai has quality problems in the generation of images, which means that the images generated by Indecent on civitai, do not currently reflect the true potential of this checkpoint.
@GTSuya_Studio Ill figure out some solutions. Takes a moment but I found best optimizations to the other Checkpoints too. If a Checkpoint can stay true to both the various diverse LoRAs intented characters, no matter how alien, and combine it with detailed realism without becoming something else, Im happy.
May be smart to add REALISM to the name so people find it easier.
Question. Im running tests on the hardest character for all realism Checkpoints: the Simpsons.
Yellow skin, weird eyes, odd face structure, every Checkpoint struggles to translate them.
Running into the same issue with Indecent now. Any prompt tips to get for example Marge Simpson to work with it?
If not, its a fun challenge for the next updates.
With or without dedicated Simpsons LoRA?
@GTSuya_Studio With! Any of them. Checkpoint wont commit to consistent strong yellow skin for example (Simpsons where even with the strongest promps possible it results in the body only becoming light yellow and oddly just the head the correct strong yellow). Also with other tests with green skin (monsters where the hands remain human, also refuses to do an obscured face, like a hooded figure). Other Simpsons factors; specific difficulty with hair, face structure and eyes. It might be caused by Civit being weird right now with their prompt changes. But had to check in.
@Strivah It will always be difficult to find a happy medium between the realism of Indecence and the unrealism of the physical appearance of the Simpsons characters: yellow skin, huge bulging eyes. For example: What exact shade of yellow will be appropriate? What exact eyes size will be appropriate?
Also it will depend how the LoRA model will be trained. If trained only on cartoon original images, LoRA model will have difficulties the generate realistic images. At the opposite, if the LoRA is trained on different image sources (original cartoon, fanart drawings, cosplay photo, etc...) the LoRA will be more flexible (not stuck to one single artstyle) and will have mode facilities to generate realistic images.
Here are some of my tries with Indecent : https://civitai.com/posts/14207391
Then you can try to play with the LoRA weight to increase or decrease realism of generated images. If you search consistency, the best is to generate a bunch of realisitic images that look good for you, and then train a new LoRA on them.
Another suggestion will be to use the semi-realism checkpoint which will have less difficuties to deal at the dame time with realistic images and unrealistic concepts.
@GTSuya_Studio Thank you for testing. Appreciate it. I might make a Real Simpsons Lora in the future. Just to fix this issue once and for all.
This one looks amazing, so I wanted to try it out, but I just started yesterday with these things.
Does this need a specific workflow or is an absolute basic flow "Load Checkpoint (VAE Decode) -> Positive/Negative promt" -> KSampler+LatentImage" enough? Are the Clips embedded or is it necessary to load them? Where to set clip to -2? I am on Comfui btw.
Edit: Looks like "ComfyUI | Lucifael | All in one Workflow" is what I might need to look at :D
No it does not need a specific workflow, but you will have to set the clip to -2. You will have to add the "Set clip last layer" node just after the "Load Checkpoint" node and before the "Positive" and "Negative" prompt nodes. Example : https://pcniki.com/wp-content/uploads/2023/07/stable-diffusion-comfyui-feature-5.webp
@GTSuya_Studio Thanks a LOT, worked ike charm :D
@Xanti It's a pleasure.
Indecent is available again!
Yes !!!!
Oh no auction system is back.
Yes... but you can still generate image with Indecent, no?
@Indecent_Studio This week yes. But next week and after? I fear the worst with this bidding stuff.
There will be so many anime and furry checkpoints getting upvoted. Indecent is growing fast, but in general people barely got a chance to even get to know it.
I just added now a recurring bid of 2500 buzz for Indecent. So as long as there will be enough users of it every day, it will stay accessible for generation.
If 2500 buzz is not enough, I will try to increase it. But I have to find the right daily value between the number of buzz earned by the model and the number of buzz I need to spend to still it availiable.
Ok... I just discover that recurrent bid is every week and not every day. I have adjusted my bid to 20 000 buzz every week. This amount was enough last time to be selected. I hope it will be okay. Sorry for the convenience. And thank you for the one fo you how add 120 buzz.
@Indecent_Studio That was me. I dont have any big funds, but I keep on promoting it because its the right thing to do. I want the best possible realism checkpoints available for everyone and after all my many many tests thats still Indecent and TAME. without a doubt, clear as day.
And many started to notice in the short time Indecent was available. It was slowly beating the usage rate of many Checkpoints that had been available far far longer.
Also I dont believe you need to put in that much each week. The first week ended at around 6000 minimal. If you go for 6-9k it should be safe and it will likely drop down fast once the novelty of the new system is over.
@Strivah Thank you for your bid. Last time with 20K buzz Indecent has finished at the 5th rank. Actually with the same amount it is actually at rank 14th, and it still have 6 days for bidding. So actually I prefere to be sure that Indecent will be selected, And in the future I do not want to check every week if thew bid if enough or not. I will see if adjusting is needed in the future.
I have re-adjusted the bid to 1500 buzz. I may be enough
Just perfect. Competing fair enough cyberealistic!
Thank you!!!!
this is really good, wish I found it sooner
Thank you !!!!!
works great, my new favourite to use!
I'm very happy that you like this model.
Works very well
Thank you !!!
Its great. It retained all of pony's knowledge. At least all characters, concepts and even styles I tried. It also works well with all Loras I tried. Though I used clip skip 2 and not -2 as in the description.
Thank you !!!!! Concerning the clip (clip skip) value, it is in fact 2 on Civitai and A111, but -2 on ComfyUI. :)
The most realistic Pony, great job, wonderfull bodies and faces!!
Thank you !!!!!
Thank you for the buzz
Any guidance on training against this model, i.e. number of steps and LR? Or should we train against Pony XL base and use with this model?
You can use the same training parameters than Pony. Personally I'm using around 5000 steps, with Prodigy scheduler (so LR set to 1).
thank you for this brother, i find this CP the most convenient. how ever, I'd love to get a good control over the lighting too. that'd be FIRE
any reccomended resolutions for this checkpoint in comfyui?
edit: incredible checkpoint btw <3
They are the same than the SDXL resolutions : 1024x1024, 1152x896, 1216x832, 1344x768 and 1536x640 (into landscape and portrait orientations).
@Indecent_Studio thanks for the responose! its working like a charm
Why doesn't faceswap work properly and even using the indicated negative prompt, a drawing similar to the GTA art continues to appear?
Faceswap do not work well on Pony based model : https://www.reddit.com/r/StableDiffusion/comments/1d678lv/face_swap_with_pony_models/
It will be better to train de LoRA of the face you want to apply and use this LoRA with pony models like Indecent.
People please make sure this Checkpoint stays available. It landed on 180th this week, nearly off the list. There's lots of overbidding on bad checkpoints going on. Lets all keep an eye on Auction, help out if needed.
Yep, I saw that since several weeks, Indecent is slightly decrease into the rank. I will try to put a little more into the Auction to avoid it to be inaccessible.
Edit : I have put 1K buzz more per week... so it will stay around the 100th rank. But if you like this checkpoint, please try to put regularly some buzz.
Indecent_Studio From what I've noticed some people have started to do last second one-ups lately. Just overbidding others by 1buzz last second to knock them down.
Bad news: Civ is limiting Checkpoints to Top 100 spots. We need to make double sure Indecent stays available. Auction will likely get out of hand with so few Checkpoints left. Keep an eye out, make sure this stays in everyone.
This should be temporary. Civ says its working on a new system that should reenable more checkpoints later.
I have increased my contribution until 15k per week... But that will be my limit. Indecent did not actually generate more than that by week. Gtsuya_Studio also contribute to 5k per week, so we are actually at minimum 20k per week. We will see how it evolve.
Limiting number of checkpoints, also mean users will only spend buzz on 100 checkpoints, instead of 200. So we could expect to receive more buzz for Indecent per week, and so maintain the contribution at 20k for a long period.
@Indecent_Studio It should only be temporary. From what Ive read from other messages from the dev team they are working on a cheaper pipeline that should increase their efficiency and return us to 200 Checkpoints. Shouldnt take more than a week if it works.
@Indecent_Studio I did my best to help out, but people are overbidding so stupid much now. Cant do more sadly. Hopefully normal Checkpoints resume next week.
The current system isn't sustainable. Only being able to use checkpoints when one pays huge amounts of buzz every week limits usage to big spenders. I injected 10k buzz last week. I can't do that forever.
@elShoggotho Thanks for helping though. The bottom was 30K this week. I tried to help as much as I could but it just got too high.
There are 3 main problems.
1-The current top 100 limit is taking longer than it should.
2-Too few people are helping with bidding. You only need like 15 people helping a little bit every week and it becomes cheaper. But barely anyone helps sadly, which goes for most models.
3-There are people who take multiple slots. Theres 1 creator who dumps 17 checkpoints every week for 30K. 17! Useless. So lame theres no rule against it.
Definitely my favorite realistic base model. I'm pretty sad it's not available for onsite use anymore. And I really don't like how the payment for yellow buzz has been , and the auction system etc.
I'm really sorry that we miss this week. But the required buzz to be now on the top 100 has doubled. It is not easy to adjust when Civitai change rules at the last moment, But, I hope Indecent will be back for generations next week. We are actually at around 30K for the bid (thank you for the contributors). This amount was enough last week to reach around position 30th. We're keeping our fingers crossed.
@Indecent_Studio thank you. I hope it does become available. The system seems to change all the time. I hope it gets better.
One of the best!
Thank you very much !!!!
Just wanted to thank you for this model. It was exactly the realism I wanted all this time; photorealistic real life detail with a high adherence to the Lora's characters.
With todays changes to the site economy I sadly won't be able to continue generating. Had thousands of ideas I wanted to work out with Indecent, but I simply can't afford it now and I don't have home-generation, so this is it. I hope the model stays around somehow, if it doesn't I fully understand. Just wanted to say thank you for these months. It was some great art to see come to life. Hopefully one day it'll be possible again.
Thank you. The situation is indeed quite worrying. I'm not sure how long I'll be able to keep the model available if most people can no longer use it on this site.
Do you have any idea which site you're going to move to now? I could possibly move the model (and all my other models) to that site.
@Indecent_Studio I wish I knew but everywhere theres the same restrictions or paid models now. You could upload to places like Seeart, but they are model stealers. I don't trust those. There might arise some new better websites following this crackdown and exodus. It's inevitable with these things and the tech gets cheaper within years. But it'll likely take a year. Everyone just need to keep an eye out what new site will pop off.
Meanwhile Civ remains the most open, they technically only broke the blue buzz model for on-site users, which in turn makes it impossible for the good models to survive Auction House. Without good Checkpoints like Indecent, theres also no point in even generating SFW. I will hang around here for some time, maybe focus on model making from now on. It'll just be kind of sparingly and random from now on.
@Indecent_Studio The website Yodayo com might start attracting a bunch of Civ users. Its a very confusing looking site but behind the weird layout seems basically the same thing as Civ. They banned NSFW a year ago but it returned again this year. They have some realism rules. But Pony Cyberreal is there too so Indecent should be fine. I done know anything more. just a site to keep in mind.
Hi I am new to the overall AI image editing workflows. I really enjoy what you have done here. Just a quick question, how can I set the output image to look more like my source photo?
I have added the pre-prompts and negative prompts. my settings are [Randomize], [Steps 25], [CFG 5], [DPM_3M], [Exponential], [Denoise 1], [Stop at Clip Layer -2]. and i get outputs that have 0 relevance to the source photo. what can i do?
Reduce the denoise value. More de denoise value will be close to 0 more the generated image will look like the source. And more de denoise value will be close to 1 more the generated image will be different from the source.
@Indecent_Studio thanks. :) i tried playing around and found that my prompts only kicks in at about 0.9 onwards.
Amazingly realistic checkpoint, which follows prompts near perfectly.
Thank you !!!!!
became my main checkpoint the instant i used it. very good prompt adherence and incredibly good at generating a variety of ethnicities, skin tones/textures while nailing both professional and amateur photography quality. easily the best realistic pony checkpoint imo.
Thank you !!!!! :)
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.