"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 š
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 ā 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:
ā
Try: masterpiece, best quality, realistic, [your subject, actions, clothing, etc].
ā Avoid: 3 paragraphs of random junk you saw in Discord
Negative prompt:
ā
Try: bad quality, worst quality, low quality, [something very specific to the prompt you're working on right now]
ā 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:
ā¦ Sampler: DPM++ 2M SDE / DPM++ 3M SDE / DPM++ SDE
ā¦ 30-40 Steps.
ā¦ CFG: 3-6 (low for more realism, high for more stylization)
ā¦ 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).
ā¦ 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:
ā¦ Sampler: Euler A
ā¦ 15 Steps.
ā¦ CFG: 11
ā¦ 512x768 resolution or whatever.
ā¦ No highres.fix/upscaling, no FaceDetailer/ADetailer
Description
FAQ
Comments (61)
we want the new lustifyš (sorry to hijack this post)
This one was in work even before I've decided to make a new lustify. Don't worry, lustify's also in training, I've been working on them in parallel
DAMN! absolutely legendary model! made some epic personal merges with this, I'll definitely be sharing a merge with this one in it in the future, and I'm gonna credit you so hard! thank you, I'm really looking forward to seeing this evolve even further, keep up the great work!
Thank you for the understanding, appreciate it!
v5 is amazing! Awesome work!
The images can come out a little artifacty - especially at mid/ long range, but I'm using https://civitai.com/models/1513492/gonzalomo-xlfluxpony to refine at a low denoise to tidy them up. So far the range of styles and people you can generate is amazing. Looking forward to exploring more!
Love your results, thanks for the review.
What kind of "artifacts" do you mean?
coyotteĀ Sort of grainy, eyes looking in different directions, low quality looking. This is without face detailer though so I may have better results with a different workflow.
prompt_bit_sorcererĀ yeah, you need to use highres fix + ADetailer/FaceDetailer. This is like a must for really quality results. I can upload my simple ComfyUI workflow where everything is already set up
Here's the link https://files.catbox.moe/147f5i.json
Agree with the comments. Love the creativity and prompt adherence. I use IntoRealism v3.4 as refiner, imho better than Gonzalomo although those are pretty respectable too.
https://civitai.com/models/1609320?modelVersionId=2028988
Its cause some amateur look LoRA is mixed in with probably default block weights. I mean amateur as ones that are supposed to do that "amateur" look, not that its poorly made.
They unfortunately also increase noise/jagged edges/pixelated look.
Could be in theory fixed by some consistency LoRAs (in theory).
MescalambaĀ what made you think I've merged any LoRAs into it? I deliberately state that it is a trained checkpoint xD
I did include the amateur set I've used for my LEAKCORE lora into the training tho, you have a point there. But it's like only 160 pics out of ~10 000, I don't really think it could affect the quality that much.
coyotteĀ After testing it quite a bit I cant find any visual problems with model, neither artefacts. Its user error, either prompt or workflow. It makes very decent pics even without HiRes fix (better with it, but that applies to basically every model due way image models work).
MescalambaĀ thank you for the feedback!
Exatly what i thought. Great adherence, top level.
When you use characters(prompt or lora), it will be anime looking.
Thanks for the review.
Yeah, the issue you mentioned is honestly one of the hardest to deal with. To really solve it with the level of precision and quality I want, itād take a lifetime. The model doesnāt default to a flat, generic anime look when you prompt for anime characters, and it hasnāt lost its character knowledge (which happens a lot with lazy merges), so honestly, I count that as a win already.
We can work around stylization a bit by using real-looking SDXL checkpoints as refiners. But fixing this kind of thing in-house is slow and expensive, and I just canāt keep dedicating this much of my life and money for free. Iām insanely grateful to the people whoāve donated and helped me cover some of the costs - but even with that, Iām still losing a shit ton of money doing all this checkpoint work.
To get rid of that anime-ish look in anime characters, Iād need to manually collect solid cosplay datasets for every character - or at least the top 500 to cover like 80% of use cases. And even then, theyād need to include different lighting conditions, different actions, different settings. On top of that, the people who edit cosplay pics aren't helping - thereās this obsession with airbrushed skin and unnatural looking faces, which makes them borderline unusable for training.
Still, Iām doing my best to improve every part of the checkpoint. Maybe Iāll figure out a better way to handle characters later. Iām not giving up, just trying to be real about the limits.
coyotteĀ So, its 100% percent illustrious. Other realistic models merge with XL(with more or less success, in every case they lose a ton adherence). Pony behave good with XL, Ilustrious not so much.
karelchima629Ā I'm retraining Illustrious weights rather than merging/destroying them with XL to preserve as many concepts as possible. So yes, it's 100% Illustrious, which is the point of this model.
coyotteĀ I mean, there are trained models(not just simple merges), which merges with XL to eliminate the anime look. But yes, its impossible to keep adherence and being photorealistic.
Is this based on Illustrious or NoobAi? Since each one has a different character database.
It's a chimera. I've made a merge of Illustrious UNET+ NoobAI v1.1 epsilon pred CLIP, then trained on that
Judging from the displayed pictures, it seems more suitable for drawing rich special effects. I can completely replace those weird special effects LORAs with this model to free up my hard drive.
For me, pony damn v2 still your best creation, the hard realistic image combined with one of the best prompt adherence model I've ever seen so far, it's a rare gem since it can do 2d to photo realistic image so well, even works well with tons of loras.
is that version gone from the site? :/
Could you point me to the checkpoint you mention please? =)
That model doesn't seem to be available anywhere on the internet anymore. If someone has it, PLEASE upload it somewhere for the love of god
@mark_sausage439Ā @nullĀ yes sadly it gone from everywhere, i still have it, but i don't think the author wanted to shares it anyway, so i respect them.
The realism is impresive, the bad thing is that sometimes it follows the prompt and sometimes donĀ“t. Merging it with other illustrious model fixes that but makes face generic again.
the model really makes you go "damn, that's pretty good"
And can you theoretically merge the LUSTIFY and DAMN?
DAMN is really good for SFW, but loses in NSFW to LUSTIFY
do you recommend Lustify as second pass?
it's always a great idea if you want to give it a bit more realism. you can use my refiner workflow for that
tested with DPM++ 4M SDE works interesting
i've only seen an extension for forge mentioning this, and that extension is the ONLY thing that mentions a 4m. so weird. what is it like?
Loving this checkpoint. Good job! Able to produce some really awesome clothed stuff, that I don't have to wrestle with the model to keep the clothes on.
Where I'm struggling a lot is the hands and feet, so wonky most of the time. How would you deal with this?
Also the texture of the clothes/fabric is almost always the same, cotton-like with vertical stripes.
A type of mix between this and your older Damn v2 would have been awesome. V2 had some incredible clarity that I don't think I've seen again.
Hi! I was wondering if you recommend any sources for me to learn how to create Lightning Pony Checkpoints.
I'm trying to build my own unrelated Pony checkpoint and I want it to have a Lightning version as well, but I don't know where to start.
Hello, fellow Wile E. Coyote enjoyer.
It just merging lightning lora into the checkpoint. That's really it
@coyotte Hello there! It's always nice to see a looney fellow in the wild!Ā
Oh, gotcha! Didn't know it was a merging process, thanks for pointing that!
Is there a specific Lightning lora for pony, or the standard SDXL works just fine? And if you have the time, could you point me to any source of information on how to do a merge?
@Harriet_hĀ it's just a standard SDXL lightning lora, either 4 step, or 8 step ones. You can merge lora into the checkpoint either by running pipeline "Load Checkpoint" -> "Load Lora" -> "Save checkpoint" in ComfyUI, or by using kohya_ss
@coyotteĀ my brother I was expecting for the merging process to be a waaay convoluted and complicated, lol. Thank you so much for helping me with this!
Hey, what is the absolute lowest resolution your model looks good at?
I usually would make my images at 768x1024 for a significant speed boost and then refine or upscale after. but I noticed i get deformed low quality images at that resolution on this checkpoint, but then the quality completely changes at 832x1216.
Nevermind just discovered dmd2
I really appreciate the time you put into this!
Thank you so much
Thank you!
Could you tell me from which pack the PaletteCanvasGradient node is in your workflow for previewing generations? Comfy manager couldn't recognize it, only ComfyUI-Inpaint-CropAndStitch
None. It's my own node that I didn't release, but I will once I'm happy with it.
@coyotteĀ hi. is there a way to run this workflow without the palettecanvasgradient node?
@coyotteĀ Okay, I'll be waiting with impatience. I like the photorealism of your model, including with the IPA adapter connected. Keep it up!
@tierastronautĀ have you tried deleting it from the workflow? Like, it's not even used in 90% of the preview images
Love how you roast the people who throw whole paragraphs of random junk at the model. Another thing I see is people adding a dozen LoRAs as if that wasn't going to absolutely confuse the model. Just because a LoRA is called detail enhancer doesn't mean it's going to enhance the details... I always try without LoRAs and with simple prompts first, some people seem to front load a truck of tags and LoRAs like it was a ceremonial ritual.
By far my favorite realistic illustrious/pony checkpoint and it's not even close.
I got an idea (probably a really obvious idea to others), from the Coyotte Refiner workflow, to generate images iwth damn and then upscale them with lustify to get even better realistic textures. The results are insanely good!
Thanks for the time you've put into this!
Praise this usage guide, lol. It should be stickied everywhere.
This is the best realistic IL model and it isn't even close. Will there by any more updates?
This model produces amazing results, very unique as well
V5 is great šš»šš»šš»
Trigger Words: DO NOT USE SCORE_ TAGS!!! AAAH
Works quite well once i've worked out the kinks on my end.
How well does this work with DMD2 LoRA?
Best realistic Illu full stop!
Does this work with my own character lora?
Details
Files
damnIllustriousPony_v50IllustriousNoobai.safetensors
Mirrors
damnIllustriousPony_v50IllustriousNoobai.safetensors
damnPonyIllustrious_v50IllustriousNoobai.safetensors
damnIllustriousPony_v50IllustriousNoobai.safetensors
damnIllustriousPony_v50IllustriousNoobai.safetensors
damnIllustriousPony_v50IllustriousNoobai.safetensors
damnIllustriousPony_v50IllustriousNoobai.safetensors
damnIllustriousPony_v50IllustriousNoobai.safetensors
ILLUSTRIOUS_REALISTIC_damnIllustriousPony_v50IllustriousNoobai.safetensors
damnIllustriousPony_v50IllustriousNoobai.safetensors
damnIllustriousPony_v50IllustriousNoobai.safetensors
fb-c17.safetensors
damnIllustriousPony_v50IllustriousNoobai.safetensors
damnIllustriousPony_v50IllustriousNoobai.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.