"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
Lightning version.
Recomended parameters:
DPM++ SDE
Karras / Exponential scheduler
8 Steps
Cfg: 1-2.
Highres.fix: upscale by 1.4, 4 steps, denoising ~0.4, upscale model "nmkd siax" or "nmkd superscale"
Due the nature of "lightning" models, this version might follow your prompts worse then the regular one. Also, I've noticed hat the lighting type tags need more weight in this version, for example "(neon lighting:1.3), (cinematic lighting:1.1)".
FAQ
Comments (47)
Kick Ass. ADetailer / Hires for sure on faces (no big deal), but wow you nailed it with v3.
Scores... pheh.
Thank you! Yeah, I designed this checkpoint to fit my workflow, which always includes adetailer and highres fix for the pics I Iike. It's very opinionated when it comes to stuff like that.
Honestly at this point scores aren't necessary. I use them for some pics to get a 'different' aesthetic
My experience with your model is that the Lightning version so far works best with Euler a, not DPM++ SDE. Same with the Pony version, although the difference isn't as much there. Almost anything else than Euler a gives a choppy image (also before hires)
That's weird, because you can clearly see that I'm using the said parameters and have no problems. Can you upload some examples of the broken sde outputs? I'd like to investigate that
I had the same experience. But for me this is the case for most of the Pony models. I could be an automatic1111 problem. I don't know.
And then you have to choose the schedule type for sampler method which has to be Karras, right? For me Euler a works perfectly with Automatic. The other samplers doesn't.
@coyotte Came back to the model, and funny enough I don't have the same issue with DPM++ SDE any more. Could be because of a change in Forge.
I keep dreaming that this model will become available on Civit generator.
Haha, me too
Okay, I made a silly post when this model came out but in all seriousness, it's a great model! For what I use it for, it does better than Pony_realism which is the most popular. Cyberrealistic_pony is on par but I find myself using this model more lately. Great work! Keep it up and I can't wait to see your work on Flux when it becomes popular!
I liked that 'silly post', just didn't know what to answer xD
I would really like to work on flux, but right now it'll take much more than a single RTX 3090ti I have. We'll see.
Thank you, glad you're enjoying it!
I've created an article on "Refining Pony Models with SDXL/1.5 for Enhanced Realism". I think a lot of you guys will find it interesting and useful. Please, give it a read https://civitai.com/articles/6651
is this compatible on Fooocus as a refiner?
Very nice work. However, I hope you aren't ashamed of your V2 work. It was a really great version. I would say V3 is a great realistic model, while V2 gives me great 2.5D results if my goal isn't complete realism. Thank you for all your work.
Thank you very much!
I am not ashamed of V1 and V2. It's just... I can be very picky about some things. And those versions didn't handle them well.
V3.1 or V4 (not sure how to name it yet) will be out soon. It'll probably be the last one in the "DAMN" series, as it finally does what I was trying to accomplish.
I actually agree on this. Your V2 is my favorite of all your versions.
@coyotte please consider bringing back V2 for those that enjoyed the model in different ways, while 3 is nice, i get some finer results with V2.
V3.1 or V4 (not sure how to name it yet) will be out soon. It'll probably be the last one in the "DAMN" series, as it finally somewhat does what I was trying to accomplish.
I got this model through your fantastic workflow, and it is great, very responsive to minimalistic prompts, and fast. Needs some after detailers and stuff, as do all Pony models, but i got all that in your workflow. No wonky hands and feet, so far, great job! Thank you.
First of all, this checkpoint is amazing and I can't thank you enough!
But I want to ask you if you have any tips about training Loras for this model in specific.
I tried to train a celeb likeness using DAMN as base, but it didn't come out good :/
I have noticed that with almost all pony derived realistic models that they can pretty much only do slight variations of a single face. This is not a critique of this model at all, I am just curious if anyone knows why that is?
Known problem of pony. I think it is because the pony base was heavily trained on anime/hentai pictures and these does have s specific aesthetic and are not very detailed (because they are drawings after all and these concentrate on the key facial features). Less details = less variance = less different faces.
There's a few things you can do. You can use negative prompts to discourage the "default face", so if you have a default Asian face, try making explicit prompts about the face, like, instead of:
Positive: Beautiful young woman, Ballroom, Dancing
Negative: ugly
Do:
Positive: red eyes with light makeup, happy freckled young Irish woman, broad smile with vampire teeth, Ballroom, Dancing
Negative: (Asian:0.1), horns, ugly
You can also use a face detailer with a different model entirely. I love the "better than words" checkpoint for this. Just Inpaint a face you want.
@turiyag thx, both suggestions make a lot of sense!
This has probably been my favourite model. Was my goto for a long time. Is it, like, not usable anymore on CivitAI? Doesn't seem to show up in search results, and the run button just brings up a list of services it's "unavailable" on.
can this checkpoint make(good) footjobs?
Damn, what a beautiful checkpoint
I'm gonna be honest, I don't see much of a difference at all between facial generation in standard pony and damn. Maybe pony has just gotten better? I'm doing side by sides with with both checkpoints and although they're different, they're both incredible.
Wait, Pony v6 is realistic out of stock, or i didnt understand good.
Its abandoned and hated by its creator! Its top 3 all time, but thats the situation.
Can you imagine to be hated by your creator? Yes, we can imagine it.
I'm sorry, but what's the point of that comment? I don't hate the community, you, or the mythical "we" you're talking about, just the realistic pony checkpoints in general, because they all have a couple of things in common and there's nothing that can be done to improve it.
These fucking gigachad-like faces with huge prominent jawlines, these dumb-looking default facial expressions. These bodies that look like they are made of wax (I'm not talking about skin texture. I mean how the bodies look)
@coyotte I was sad, and it was just a dumb poetry.
What you described about faces and bodies, jaws. Are you talking about Flux? Pony doesnt have this. Facial expressions are, why i moved to Pony from SDXL. SDXL is weak and puny, you must weight complex expressions to the sky and the result still shit. With pony you can make complex emotions with 1 tag.
Good luck for your next project.
@kristansharon558 Pony was like that for me initially as well. Then something clicked, and now I can't unsee what I've highlighted in the previous comment.
And by faces I mean these abominations https://www.reddit.com/media?url=https%3A%2F%2Fi.redd.it%2F41hhh6fkk7pd1.jpeg%3Futm_medium%3Dandroid_app%26utm_source%3Dshare
Of course, it's not as extreme nowadays with the newest pony checkpoints, but it's still there and I can't unsee it. Every "Pony face" looks like that to me. Every "pony body" looks like it's made of concrete/hard wax, not something soft. This is just my experience and I don't say that it's the truth.
@coyotte Same me, just with flux.
I don't know why or how, but this model has some kind of intelligence others don't have. This isn't an allusion to something else. It's not that kind of comment. Damn is just incredibly interesting and unusual. It understands nearly everything and will often 'comment' in some visual way, which is often genuinely funny.
An alternative explanation is that I have finally gone insane. It was going to happen eventually.
Was there even a damn! v2? Looks like it's not available any more?
@coyotte is there any chance of Damn v2 being available?
I've used this model quite a bit, including damnv2 and I think there is a substantial uplift in quality especially at the higher token counts with more complex prompts. I understand the want but IMHO with 15k or so generations on this model you won't really see a crazy difference positively.
Still top2. If the retired old man made a v4, it would be probably no1.
Happy new year!
what do you think is number 1?
I would also like to know what you believe is number 1
Cyber is no1, just for the best balance between realism and flexibility, booru adherence
"old man" bruv I'm not even 30 😭
have i got news for you then :)
can you petiton the mods to make this checkpoint active on the generator? I can't believe it's not one of the chosen 200. I maybe f with 2 out of the 200, and this checkpoint is on of my mains at Tensor.