"DAMN!" (c) - my first thought when the finetuning ended and I saw the final results.
BEFORE USING THE MODEL, READ THE DESCRIPTION, SPECIFICALLY THE PROMPTING AND PARAMETERS SECTIONS.
If you want to support me or help me cover the costs:
https://boosty.to/coyotte/donate or Bitcoin 1CoMZvrXtRXbz7BCzdvxifSjHqhaf1yUZs.
Every donation will go towards developing chekpoints, workflows and loras (this stuff is costly). I have big plans, but little means to fulfill them 💀
YOU ARE NOT ALLOWED TO USE THIS MODEL AS A PART OF YOUR MERGE, UNLESS YOU CREDIT ME IN THE DESCRIPTION.
Important prompt structure for Illustrious-based version.
Rule #1: More tags ≠ Better output. It's actually other way around. Do you really need 89 tags?
Rule #2: Use danbooru tags.
Rule #3: If your prompt causes physical nausea to read - fix it.
Rule #4: DON'T use Pony loras on Illustrious.
Positive prompt:
✅ Try: masterpiece, best quality, realistic, [your subject, actions, clothing, etc].
❌ Avoid: 3 paragraphs of random junk you saw in Discord
Negative prompt:
✅ Try: bad quality, worst quality, low quality, [something very specific to the prompt you're working on right now]
❌ Avoid: dozens of redundant anatomy tags like "missing limbs, bad eyes, floating limbs, etc", negative embeddings made for specific checkpoints
Pony's quality tags like "score_9, score_8_up, score_7_up" AREN'T doing anything on it.
If you choose to ignore the guide and are getting trash results - please also ignore the "rate" button.
My simple yet organized workflow I've used for creating the preview samples: https://files.catbox.moe/147f5i.json
RECOMENDED PARAMETERS FOR BEST RESULTS:
◦ Sampler: DPM++ 2M SDE / DPM++ 3M SDE / DPM++ SDE
◦ 30-40 Steps.
◦ CFG: 3-6 (low for more realism, high for more stylization)
◦ The model was on trained mostly on 1024x1024, 832x1216, 1216x832, 896x1152, 1152x896 resolutions, so it could help the quality if you used these (although that's not necessary - it behaves nicely on other aspect ratios as well).
◦ Highres fix: Upscale by 1.35-1.5, denoising strength ~0.4, upscaler - 4x NMKD Superscale or Siax.
For the distant shots of people you probably wanna use highres fix/upscaling and/or adetailer.
PARAMETERS FOR WORST POSSIBLE RESULTS:
◦ Sampler: Euler A
◦ 15 Steps.
◦ CFG: 11
◦ 512x768 resolution or whatever.
◦ No highres.fix/upscaling, no FaceDetailer/ADetailer
Description
The first upload.
FAQ
Comments (139)
Thank you for you amazing work! I really love the skin texture the model provides, also it is pretty flexible even with an extreme tagging style. Definitely deserves thumbs up!
Thank you! Stay tuned, more is coming
Damn, this looks promising. Downloading now. Do you have any recommendations regarding the source_ or rating_ tags that pony uses? Should we use them here?
Thanks a lot! Would really appreciate it if you let me know what you think about the model after you give it a test.
Regarding the "source_" tags - If the only thing you want to gen is realistic images, than no, I don't think you need to use that. However, "source_furry" or "source_pony" can probably give interesting results, untested.
"rating_" may be useful to direct the model away from the NSFW with the "rating_safe" tag or whatever it was in the original.
Ok, the skin texture is by far the best, better than Vivid's and Vendo's.
The backgrounds are quite varied, far more than other realistic models that are NSFW friendly, plus crowds are present, BUT...
It has it's flaws, the tag 'close up' which I often use on Vendo's and Vivid's in this model is somewhat glitched and only shows close up's of penesis (tf?).
The hands at medium shot are far far inferior to other models.
Faces at far shorts are almost unrecognizable, I'm talking shit quality here.
Stabelished characters, this might not be for everyone, but your model has trouble with Cosplay characters, and that alone can make it unnasable.
Also the faces, they seen to be like the old 1.5 models with the samefame syndrome, with Vendo's I can type the character's name and get a facial structure similar to that character even without LoRa's, this models fail miserably, it failed at generating Cynthia and Misty from pokemon, this to me is not enough to use the model.
Don't get me wrong, the realism is there, but that alone can't carry the model specially for people who are looking at making AI porn and not simple renders with just one character with their hands hidden.
Thanks for the feedback.
Huh, that's weird that you experience all of that stuff with my model. Can you post your results so that I see what's wrong? I've never had these problems with "close-up" tag.
I've tried generating cosplay pics before - had no problem with that. It was an earlier epoch of the model, but still.
To get Cynthia and Misty, you should type cynthia_\(pokemon\) or misty_\(pokemon\) since pony uses danbooru tags. I'm getting great results from all the character tags I've used so far. It even worked for lusamine as strange as her hair is. Also make sure you use adetailer for non closeup shots. Easily the best realistic pony model I've used so far.
wow this looks incredible, legit game-changing for a PonyXL model, by far the best iv seen. Hoping lora's work fine aswel as the Pny's understanding of tags, top work dude DAMN! was my reaction looking at those preview pics. Downloading rn!
Holy shit.
Can this be made available for use on-site? I have to try this out but I'm definitely tech-illiterate.
It is enabled. It's just civitai doesn't allow the on-site generation for this model for some reason. I'm talking to the support to fix it
The response from the support - "Ah OK - so checkpoints are a limitation for us at the moment, we only able to make a certain number available on the generator, so it may be quite a wait before we're able to scale to include all of them."
@coyotte damn that sucks. I'll anxiously wait for it to float to the top
This model does look good. Better texture and colors. However, the other real models read prompts better. I think the merge you're trying to make left out too much of the pony side in order to achieve the look you have. It's close, really close but I don't think it's quite there yet. I may try and merge it with another pony model for better prompts. Keep working on it!
Overall a good model.
I was having a hard time getting faces to look normal. Even after Adetailer and low denoising they end up a little soft looking and "samey". Hand's seem a little eldritch, too. I think I've found settings that minimize these issues.
Thank you for the feedback!
Yep, the problem with the faces will be fixed in the next updates.
incredible, the best Real pony model i must say!
Thank you for your kind words!
Thanks for the feedback y'all. I have taken note of the issues you mentioned, and will post another version to fix them when it's ready.
Pics look great, but I have not been able to get a "normal" result using a character LoRA. Any tips or hints on this?
which lora were you using?
Sorry for not answering yet, was afk some days... I used a selfgenerated pony character model, realistic person, generated onsite here.
@Fennek_ i see. well i'm gonna upload a new version soon, it should be more compatible with pony loras.
First realistic pony model that achieve good skin textures! Good job. Looking forward to updates.
Pretty great model. Faces are a LITTLE lacking with Loras, but everything else is pretty spectacular so far. You really nailed it on the textures. Thanks!
Thank you!
Which loras were you trying to use? I'd like to see it for myself, so that I undestand what is wrong and why it is that way
@coyotte My own that are trained on base PDXL, which I assume is the issue... eyes are always a little distorted and unless it's REALLY up close, details are fuzzy. I'm using Fooocus too, so I'm not kicking in ADetailer or anything like that. It's pretty par-for-the-course I've noticed with all of these new PDXL merge checkpoints except Everclear... that checkpoint seems to always work well.
Either way, not a big deal. Just thought I'd mention.
@r0bman @r0bman thanks for the feedback, it's really important for me to make the model as compatible as possible with the existing pony ecosystem. I'll see what I can do in the next version
Best realistic model I've seen thus far, and one of the only ones that retains the ability to make femboys (probably because it's a trained model and not a merge). Can't wait to see where this goes.
Oh, actually you just gave me an idea to include femboys into the dataset. Thanks!
by far the best realistic pony model, it's awesome! the main issue I see is that the model seems to be very biased towards make-up. every girl automatically wears a lot of make-up. its impossible to get natural looking faces without adetailer.
Thank you for the feedback, will try to fix that in the next version
@coyotte nice, can't wait for an even better version!
Decent model but as others have mentioned it does struggle with prompt instructions and after merging it with my own Pony and SDXL LoRA model (both kinds work on that merge) I noticed a steep decline in promptability as well as the faces included in you training data much have been somewhat repeated because rather than just loosing minor compatibility with SDXL character LoRA it all was completely gone and every attempt at previously working LoRA returned the same face which had kind of a slight uwu smile and distinct cheek bones and very super model like. You sure you didn't merge in face focused LoRA just to try and fix face issues? Cause I understand if you did but it is not a good approach.
Thanks for the feedback!
I am aware of the problem with the faces and already working on it. The last iteration messed up faces and I didn't notice it in the testing phase.
The loras don't really work with this model for the same reason SDXL loras don't work with the base Pony - a lot of finetuning have been done and the weights have shifted drastically. And it's probably gonna get worse as I continue training. I have an idea how to reduce the damage and make them more compatible, gonna test it in the new version.
@coyotte I understand you. I can help you out a bit. I already extrcted decomposed weights as LoRA from model because I wanted to investigate solution for faces anyway. So gonna run some LoRA block control generations with my hybrid model that does the thing you want yours to do. So it will be easier to gauge weight location. You can add me on discord if you want. check my profile under my user name and join date for link
The model is a bit wonky on outputs, lots of warped bodies and mangled fingers, but overall still good. The backgrounds aren't great but the lighting is very good, lighting is one of the most important factors in the appearance of realism. On the rare occasion, the outputs are 100% awesome.
Can you please give me some examples of your generations? Upload them to catbox or something. Because the model shoudn't output warped bodies, so something's wrong on your side maybe?
@coyotte Apologies, I didn't keep any of the failed generations, I've got a handful of nice successful ones. I'll check in again if/when there is an update to the model :)
@coyotte Here's a few of the successes
https://files.catbox.moe/g77t7j.png
https://files.catbox.moe/u6k80h.png
https://files.catbox.moe/ncraj1.png
@VelexiaOmbra really nice results!
i'm working on a new version that should fix all the problems, will be released soon
For me, this model is really bad at medium/distant shots.
are you using highres fix?
This is the most realistic pony I've seen, good job!
I will test it a bit more and bring some useful feedback.
thanks a lot, looking forward for it!
So good. Finally a model with SDXL image quality & Pony capabilities. Please ignore the criticism (on a very high level) here & download immediately!
thanks a lot for your kind words!
Yes this is great, my only wish is to create darker images, all seem to be sunlit.
Add ((Night, Dark Background)), to your prompt. Those should do the trick in no time! :)
Danbooru (the dataset Pony was trained on (as far as I understand it) knows "night" specifically.) You can browse ALL the keywords it should know here:
https://danbooru.donmai.us/wiki_pages/tag_groups
* not all of Danbooru's keywords "work" with pony, it depends how much training data it was fed for each. If you use the "correct" words for things in your prompt it works MUCH better than just guessing. "Big Boob" will work, but "Large Breasts" is directly in the dataset, for example.
My generations keep getting things on their face, like they want tattoos on their face or something. Tried messing with all different settings.
New version ETA: May 15-16th. Actually, two versions at once. One of them is... interesting, to say the least.
oh this is gonna be fun
this new version is not yet 100% realistic - more like 85-96% ...
so, the question is: have you tried some additional LORAs to add the true photorealism/photographic effect to it?
I did try to use these loras, but they aren't making much difference. It's still not on juggernaut's or other sdxl checkpoints level of realism. I mean, that's not surprising, pony is 100% cartoon model, so it'll take some time to get somewhere close to juggernaut's level.
@coyotte https://civitai.com/models/299086/epicphotoxl-lora - try this one, as well as some other ones, that were trained on original Pony, they DO add some more photolook in my tests
@OliviaRossi interesting, i will give it a test, thanks a lot
@coyotte Also boring-reality LoRA's: https://civitai.com/models/310571?modelVersionId=348837
they add great aspects when used with my merge. Will also try around a bit on DAMN, thanks for the great model!
Got My Recipe for datass rev 2 in my model description, maybe you can take something from there.
@datass280 thank you!
@coyotte also - do a sampler test (all available sampler + scheduler combos) = they are also very much change the perception, and --- negative prompts, all of this can "push" model to more photorealistic look
@OliviaRossi except epicphoto lora does not accept that we share merges ;)
@mskiller51 yes, but at least it can be in recommended resources section, as well as best positive+negative prompt with sampler/scheduler combo = it all drives generation either into anime, or realism
My favorite LORAS to boost realism (all specifically trained for PONY):
https://civitai.com/models/437039
https://civitai.com/models/454172
https://civitai.com/models/450852
https://civitai.com/models/327426
Textural embeddings:
https://civitai.com/models/390305
https://civitai.com/models/411532
I haven't tested them too much with this Checkpoint, but they tend to play well with all Pony checkpoints. Start them around 0.5 and tweak from there.
I use PONY because the underlying dataset is insanely well trained (compared to base SDXL), so it understands prompts in much less frustrating ways. It's a cartoon-based one so getting 100% realistic is not, um, realistic. Plus you can refer to the https://danbooru.donmai.us/wiki_pages/tag_groups tags so you can get ideas what words to use for concepts (that's only true for PONY based Checkpoints.)
Also, if you want naughty stuff, add (rating_explicit) or (rating_questionable), "NSFW" is pointless. (Source_Western) indicates not to use as much cartoon and anime sources. Source_photo sometimes works but not reliably.
@Gouda RambleStyle01 and pony_negative for PianoMix were unknown for me - TNX!!!
It's also worth trying to not use "score_9, score_8_up, etc" prompt and not use negatives. This results in that amateur realisitc aesthetic. I've uploaded some examples.
@coyotte You sell yourself short! I find this model about the same realism as Juggernaut. I use a few models that are more realistic, but this is the best PonyXL derivative.
@RedAstronaut thanks, that means a lot to me!
Absolutely stunning! good job!
thanks for sharing this great model, keep it up 🌹
Thanks a lot for your kind words, hope you'll enjoy the model!
You levelled up! Congrats on this amazing model. This is an excellent checkpoint.
Thanks, really appreciate that!
Damn Pony!!!
I was banging my head up against this model for the last couple of hours, I had a few decent gens but most of them were garbled, i should of went to bed when the new version dropped, but i finally figured it out. Disabling/bypassing my unused lora loaders in comfyui finally let me consistently generate to great images, hallelujah!
Glad you were able to figure it out
DAMN! 5 stars.
Edit: Why do you say in the details not to use Eruler A?
Thanks!
Euler A is known to produce soft-looking images with basically no skin texture. And the whole point of realism and this model in general is it's textures and details. So if you do want to use Euler A for some reason, make sure to use different sampler in your highres.fix pass (you might need to enable this option in settings tho).
Gonna take a small break from training "DAMN!", like for a couple of weeks. I've completely exhausted myself while working on the v2 :( I'm not abandoning it, don't worry.
I might release some fixes or even experimental merges (shoutout to @S1LV3RC01N for telling me about DARE merging), but mostly will be preparing for SD 3 training.
I hope you guys understand 🙏
did you unlist sweetspot for a reason?
@civitai956 i've made a couple of test myself and turned out "sweetspot" got the worst out of both "subtle" and "extreme". Ironic, but it just the way it is. Do you need it?
@coyotte no i just had it downloaded, thanks for the update!
HI, you removed subtle and sweetspot ? are they redundant ?
@MoebiuS451 hi. "Sweetspot" was removed for the reason i've mentioned above, "subtle" was removed because no one was downloading it, so I figured people didn't like the flat style it had.
The update is fantastic so far. Faces with Loras in faraway situations are looking quite better. Up close they're perfect. Thanks!
Glad it got better for your case of use, enjoy!
The images are super cherry-picked. The actual results are very cartoon.
huh? that's a very bold statement without a demonstration, upload your examples. you're probably doing something VERY wrong, unless we have different definitions of "very cartoon"
Here's a batch of 6 photos, proving that cherry-picking isn't needed. Idk, does it look "very cartoon" to anyone?
https://files.catbox.moe/r1zbwi.png
I had these same issues, until I learned to prompt better...
Now, it's freaking amazing!
You made an amazing model, my man.
@blocker3d thanks a lot for your kind words
@coyotte Thank you for your time and effort on this!
@shadow0 Output is a little cartoony, but definitely not very cartoony. But the samples are not cherry-picked: the samples are a bit cartoony! more so than my results even! So I think the samples are honest and representative. Check out the Gallery, some more realistic than the samples.
This is the most realistic NSFW model I've found. Top notch checkpoint, one of the best.
Nice model, good realism, but I think the images generated are overexposed.
have you tried different samplers? "2m sde heun" makes the images with a lot of contrast. also cfg has a big impact in that
here's a low-effort prompt grid that shows the impact of the said parameters. it has metadata so you can throw it into your webui https://files.catbox.moe/glj9f5.png
also, worth trying to minimize the usage of "score_9, score_8_up, score_7_up, score_6_up" prompt as this sequence pushes the model back to it's cartoon roots and is kinda overtrained in base pony. here's a demonstration with only a "score_9" quality tag: https://files.catbox.moe/h3br7y.png (warning, exposed nipples. don't open at work or in public).
let me know if you have any other questions, would be glad to help
This is definitely a solvable issue. Disable your VAE, change your sampler, and modify the CFG.
Nice model but ugly faces
I had this issue but was able to resolve it. Some tips. "gorgeous face" in prompt. Sometimes a negative or positive tag will cause them to appear ugly. Also using bad resolutions or samplers can make them ugly too. Make sure you do 832x1216, DPM++ 2M Karras, 30 steps, CFG 6.5. Some situations you may need to use ADetailer too (depending).
Try copying some samples and work back from them.
@RedAstronaut great tips
Any news on if this will come to CivitAI Generator?
I wish I knew. Civit allows it when it meets their criteria which isn't actually known to anyone.
Probably a popularity contest to conserve resources.
Thanks for answering! Best of luck getting it on the generator!
@ponyboi thanks a lot 🙏
@coyotte Though we have lots of pony real models now, this one still high on my wishlist for on-site generator. Keep going! Maybe make a new version. If I get a new computer, I'll make generations with this one, but DAMN! Can't wait for on-site.
Great job on this one. It's my favorite checkpoint now for a while. It does fun things, while also remaining close to realistic!
Edit: Just read your entire post. Can't wait for next iteration with some improvements.
Thanks a lot, glad to hear that!
Yeah, I will try to imporve the checkpoint when I get some real life stuff sorted out
Results are too bright with blown highlights. You can see this in most of the example images, including those from Coyotte. Even in rainy night scenes, characters look like they are in direct sunlight with blown highlights. "Blown highlights" means that the brightest parts lack details. Results look like photographs made by a very bad photographer.
have you read the description? i specifically adressed this question there
if not - go ahead and read it. i write this stuff for a reason.
I used my web browser's search function to determine that the phrase, "blown highlights", does not appear in your description. The fact that some of your own images display too much blown highlights is proof that you don't understand or care about the problem.
@nizur then you're obviously blind or have too short of an attention span to read a description worth 2 minutes of reading. not trying to be rude, but that's just the way it is
@nizur just move on to different model then
hello, thank ur for this, could u please make a 2step lighting version of it too? us cpu users are much much slower in generation, thank u
I will say, this model is now my preferred use for 3d esque generations. V1 was a bit rough but extreme is performing very well with just plugging in a lot of pony loras. nice job!
To be honest, I don't see myself training this model further. It takes too much time for too little results. I mean, the model gets realISTIC, but definitely not real. Pony is a cartoon model after all. So yeah, this project is on indefinite hiatus, I've changed the licensing so y'all can use it in your merges now (not that anyone cares about licenses - I've tested a couple of realistic pony models from civit and most of them seem to be using my model to some extent. I wish people actually gave credits, but what do I know). Things might change with Pony v7.
Thanks for being there guys.
Congrats, you definitelly created the most realistic and most creative Pony-based model so far. I like v2 even better than v1. I understand that it is difficult to improve upon. Maybe you can do a v3 in future after some other Pony model was improved & you can build on that.
@SubtleShader thank you!
After a couple of new versions that I didn't release, I came to conclusion that there's no point in retraining this (v6) pony version to realism - the creator made a couple of crucial mistakes in training it that he himself also admitted. The whole score_9 prompt thing wasn't a good decision and people keep using it on realistic models (which pushes the gens back to cartoon style). But at the same time, without it the gens can get messy because of all the low-quality data in pony's dataset. If anything, the score_9 prompt should be used with a lower weight, like 0.6 or something.
SD3 comes out June 12, and I think we can expect new pony to come out this summer. I feel like the SD3 Pony will be the one to adopt realism properly.
@coyotte Looking forward to it. Thanks for the score9_up:0.6 hint. Will try. I run the damnPony output through my SD1.5 merge in img2img anyway, which produces more realism. But the more realistic the Pony output is, the more realistic the end result.
@SubtleShader been doing the same thing lately, except I'm using a realistic SDXL checkpoint as a refiner. SDXXXL works for nsfw stuff, Juggernaut for non-porn.
Thank you for all your hard work. This model is incredible and one of my favorites.
Cartoon realism is GOATed. Perhaps you didn't create the checkpoint exactly as you planned, but for some of us, it's perfect. :)
Amazing.. I've been using Pony Diffusion for 3 months, and no matter what checkpoint I tried, it didn't produce the same results as this model. How did you make it? You've made an amazing achievement. Thank you. Amazing!!
thanks.
a lot of trials and errors with finetuning it, finding the best parameters and caption/img pairs
this is the best checkpoint, this does follow all the prompts and all your warning about sampler and adding score 9, you really dont have to, I have been using this for the past 1 week and it makes me happy , more than any other pony checkpoint merge this is the ultimate one, cheers mate, I hope you come up with something else , give us a twist, cheers , two thumbs up!!!
Amazing work being done here. When will the Civit team finally listen and make this available to the Generator?
Hey there again, I see you have no intentions of updating it for a while, but I'm wondering if we could get a LORA version since Civit won't be making it available on the generator for a while. Could be really interesting and useful to train a LORA from your images to make it more versatile with other Pony checkpoints. Thank you!
I also just found that we can extract a LoRA from this using Kohya. https://github.com/kohya-ss/sd-scripts/blob/main/networks%2Fextract_lora_from_models.py
Make this available for on site generation please.
adds viginas to men when you dont ask for it.
Sounds perfect. Downloading now, lol.
@ritAIin lol different strokes for different folks :)
I guess women are more represented in training data than men. You can always ask for penis and testicles if that's your jam.
Can we get a V2 back? It was a nice model and worked with different LORAs really great, better than other photo models. I miss the v2 =(
V3 GONNA BE RELEASED SOON!
Previews are looking fcking awesome
bigasp mix? 👀
@jahkeembowdy534 no