C’est un modèle Pony. J’avais le même souci, que des résultats corrompus, jusqu’à ce que je passe à clip skip 2 et SANS controlnets. J’ai entendu dire qu’on peut utiliser controlnet, mais juste pour transférer la pose. Regarde Mistoline (mais je n’ai pas encore essayé).

Enfin, peut-être que le problème vient d'ailleurs. Je viens tout juste de commencer à utiliser les modèles Pony, donc je ne maîtrise pas encore bien leur fonctionnement.

@Aiariar Jai trouvé le soucis. je suis sur Flux. problème résolu. ;)

Use adetailer, look for the option to use a different checkpoint, under the inpainting tab. pick your favoutite sdxl, which has similar generation parameters and off you go

I don't try it yet, but search talks this model is understand 'Danbooru tags' – so, using them you can switch faces, clothes and so on.
Look at this article, Characters section: https://civitai.com/articles/6555

@ZyloO Thank you!

You need to get one picture from any user on gallery below and click on "remix" button (top left corner on image's page). And then you can generate your image with this model.

You can look on images in my profile. They all generated in ComfyUI with this model. You can load any my PNG images as workflow. P.S. I was talk about v2.2 Main. Here is one review desk for all models, wtf? I switch from v2.2 to v2.1 and see your comment again.

hi thanks for your respons @xNzX im using your workflow but still get blank image generated result, don't know why mybe vram issue, i'm using 6Gb vram

@aioffline It can't be vram issue. Low vram = slow generating. But images quality will be the same. Any other model working well for you? Something wrong in your settings. "Clip set last layer" node set to -2 (minus two)?

@xNzX yeah other model can run perfectly, but i can't generated this model with comfyui in my own pc don't know why ? i'm using your workflow with "Clip set last layer" node set to -2 but still generated result with blank image

@aioffline Make screenshot of workflow and screenshot of console (after generating attempt) - I will compare with mine.

Workflow + Custom Node

https://ibb.co.com/5Wn6GgH

https://ibb.co.com/ZdqpKDn

@aioffline Workflow look OK. Can you make copy-paste text from console window? On CMD window press Alt+Space, than choose Edit > Select all > Copy. And paste here. If this board allowed long text. If not, message me in site messenger for contact.
My generating stages from prompt: https://ibb.co/yVQcTQk

@xNzX here's my logs https://pastebin.com/a4JezPqc

@aioffline Here is our difference: after generating you load AutoencoderKL model "319.11416244506836", but my ComfyUI load "159.55708122253418". Try to google wtf this is.
From my log:
Requested to load AutoencoderKL

Loading 1 new model

loaded completely 0.0 159.55708122253418 True

I was thinking before "AutoencoderKL" it's Reactor node thing, but I reload ComfyUI, disable Reactor node for pure generation - and Autoencoder still persist. So it's pure ComfyUI stuff.

during BJ? yeah that's a pony thing not specific to this model.

@djwashout Nope, just almost every single one in most generations. Even in the OP pictures.

I would recommend using zluda

@samwyse Everytime I've tried different solutions for AMD on Windows, I always have to run in a mode for less GDDR RAM or whatever despite having 16GB. Always limited in some way. Except for running ONNX models on Amuse that is cracked to allow nsfw, ONNX models run perfectly on that.

@samwyse Been several months since I tried other methods so I just tried zluda (could of sworn I couldn't get it working before) but it is working great on ComfyUI for me once setup! Thanks for the recommendation!

"Value not in list" - say to you the name of file is different. You can GUESS what file needed - and choose it manually (on every node that tell you error). Example: your program try to find "sdxl.vae.safetensors" -- but you have "sdxl_vae.safetensors" (different name - not dot but under-line symbol in middle of file name). From your log:
"'sdxl.vae.safetensors' not in ['ae.safetensors', 'ponyRealism_V22MainVAE.safetensors', 'sdxl_vae.safetensors', 'taesd', 'taesdxl', 'taesd3']" So you need manually choose 'sdxl_vae.safetensors' for this node and this will ok.

But you don't need do load VAE ('sdxl_vae.safetensors') in this model - because model include VAE itself. I've make very basic workflow for you on this model. You can load PNG image from this post as workflow. -- https://civitai.com/posts/9419085 I've include one LoRa loader, with Lora from model's author – 'zy PonyRealism Enchancer'- https://civitai.com/models/927305/pony-realism-enhancer

Loras are everything my friend, try them !

I've use 'Reactor face restoring' node for ComfyUI - to improve faces quality. You can get Reactor nodes here – https://github.com/Gourieff/comfyui-reactor-node?tab=readme-ov-file

P.S. Got MUCH BETTER faces with FaceDetailer node! You can load my images from this post (NSFW) as workflow - https://civitai.com/posts/9892901

What error do you have in console? I've using v2.2 and it's working good.

I figured out it was the file name. even thought 2.1 and 2.2 are labled differently, stable saw them the same edition. i had to add a random letter and yeah, good now. same exact MB in size, i wonder what changes if any are in 2.2 cause 2.1 works amazing

You need to google "Pony facial expressions". Use right prompt words for getting right expression. Take a look here: https://rentry.org/corneo_facial_expressions and here https://civitai.com/articles/5492/danbooru-tags-complex-facial-expressions-for-ponyxl-autismmix

@xNzX thanks for the reply. It will sure help me for generation. But i didn't found what i search exactly, maybe with a future lora or with training a lora myself.

@onimessatsu search the danbooru wiki for facial tags: https://danbooru.donmai.us/wiki_pages/tag_groups

Use: 1girl, from below, looking down, looking at viewer, half closed eyes

and mix the emoitions tags: disgusted, annoyed, plus whatever works. I couldnt replicate this exact look, specially with full teeth, haha

@salvatorefalon428 thx for reply. The sources will be sure usefull. I know i'm new with prompts and things like that and i will step up with that !

v2.1 is still here. You need to click on 'banner button' on top of page (button above example pictures). Here is direct link on v2.1 page – https://civitai.com/models/372465?modelVersionId=534642

ok got confused. So 21Main is the same as the old 21VAE?

@djwashout Yes, this is names of the same model.

In my profile you can take a link on very basic ComfyUI workflow for PonyRealism v2.2.

Look at this article "Pony Realism Compedium" – https://civitai.com/articles/6621 and in bottom line you will see link on article about LoRa creating.

Thank you, I do not think it would be so incomprehensible, I try but I can’t install Kohya, I leave for now.

For dark scenes you must using LoRa with dark style. I personally suggest Concept Art DarkSide - https://civitai.com/models/264290?modelVersionId=446346
Add this Lora and words "concept art, realistic, dark theme, low light" at start of your prompt.

P.S. Ah, you using LoRa yet. OK. Try this, anyway.

@xNzX The whole 2.2 is about darkness and blur, if you use default settings.

Can you provide some examples of 'good' and 'funny' eyes?

@xNzX Yeah let me try here soon, ill get some up here

Try to use use adetailer face or maybe adetailer for eyes.

I am trying now, yes maybe I will learn to use add detailer

trying to upload one now

https://civitai.com/posts/10391111

Separate file VAE not needed for this model - VAE yet included inside in this model.

@xNzX okay, thank you

Try this in prompt: score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, best quality, detailed, detailed body, detailed face, extremely detailed, ultra detailed, masterpiece, photorealistic, cinematic, 1girl, back shot, view from back, standing, looking back, looking at viewer

Don't know much, but may be my ComfyUI images - (NSFW) - do something as samplers. For me, most effectively faces fix is using ReActor nodes: 'RestoreFace' node or 'FastFaceSwap' node. But this sometimes need adding Photoshop manually editing - for remove from image unnecessary sharped 'restored' other faces.

Negative: low quality, worst quality, render

Positive: absurdres, detailed eyes, highly detailed eyes etc

Use Karras and Beta scheduler, where you can.

Anyway every model struggling with distant faces, because of less pixels. Use normal resolutions: 832x1216, 896x1152, 1024x1024

Problem is not model, its the damn comfyUI. Over hyped and over complicated mess with short lasting node support. Go back to 1111 or move to forge.

I found out that hyper (cfg version) basically cannot be used with this model after much testing. It breaks limbs, feet, fingers, etc. I haven't tried the more gigantic resolutions yet because on 2080ti 22g, they take too long. I need an image out under 10s, preferably less than 5 or 6. What has worked is using Automatic CFG "warp drive" and maybe Sage Attention node. Then you can use 20 steps and it will be quick. I like the look it gives so didn't want to give up on it.

So... Are you just throwing your prompt into a CLIP node; Or are you using the Advanced CLIP node..?

* With the Advanced CLIP node you can change the weight_interpretation to A1111.

Also, if you're not using the Efficient Loader Node (Which I'm guessing you're not), you need to set CLIP Skip to -2. You can do that with the 'CLIP Set Last Layer' node.

My advice:

1. Reso = 896x1152

2. Sampler = DPM++_3M_SDE (You could use the DPM++_3M_SDE/gpu too but... meh)

3. Scheduler = Exponential, AYS SDXL, or Euler (I prefer Expo but the other 2 aren't bad at all)

4. At least 60 steps (if Regular Full Model is being used)

5. Run the image through some aDetailer nodes because you're never gonna get a perfect picture out of a model.

I moved to ComfyUI almost a year ago now... and I'll never go back to A1111. The resource allocation is 1000x better in ComfyUI and anything A1111 can do; ComfyUI can do better. (My Opinion)

EDIT: I saw your follow-up post and see you're using a 2080Ti and you want to gen in 10 sec per gen. I'd seriously consider upgrading your card. 30 series are a pretty good bang for you buck.

@DogFace I thought clip skip isn't a thing for SDXL. I will try some of those nodes and see if they help like automatic CFG did. I have other cards besides the 2080ti 22gb but they are being used to run llms. Image gen falls to this one, which is why I want to optimize.

@Gore_Man SDXL and PDXL are, mostly, inherently -2 CLIP Skip. But that doesn't mean the CLIP nodes you're using know that they are supposed to use it. In fact, most CLIP nodes will default to -1 CLIP Skip.

It should go: Checkpoint Loader > CLIP Set Last Layer > CLIP Conditioning > KSampler.

If you're using LoRAs you would go: Checkpoint Loader > LoRA Loader (with CLIP) > CLIP Last Layer > CLIP Conditioning > KSampler.

The Efficient Loader makes a lot of this easier/efficient because you have the Model/Checkpoint, CLIP, LoRA, Prompts, and Weight_Interpretation all in one node. You can add a LoRA stack to it as well for multi-LoRA support.

@DogFace I am using rgthree power lora as it's the only node I know of that pulls them from the prompt like A111 did. I can tell you for sure that clip skip is correct because if I set -1 I get black image, if I set -3 I get way less detailed image.

edit: setting -4 is a cool effect.

I'm guessing the licensing, technical limitations, etc, (FLUX, SD3.5) is potentially too restrictive? I was never really a fan of Pony Diffusion V6 XL, but I heard what he had to say with regards to V7, & going the Illustrious route also.

My money is on your new Illustrious model being better though ;) Best of luck with it <3

You can't avoid face distortion with any prompt. It's technology implemented bad quality. Only fix face working nowadays. If you use external generator, use ADetailer (face fix) for A1111, or ReActor FaceRestore node for ComfyUI. In Civitai generator use generator with 'face-fix' enabled option. But even with all those features enabled – it's not always working well. We need just wait for another level of technology.

Try "comic" or "comic story", then "spoken word" / "speech bubbles" for dialogue.

@Stardeaf thanks

LOL! Feel your pain bro. On top of that, breast sizes are always bigger than you expect right?

You can try HotshotXL or AnimateDiff SDXL... since Pony is based off SDXL.

@colliecollie Hoyspot does work but it has a limit of 8 frames and 8 fps, if you put more it gives an error and you have to restart the PC because it seems to freeze.

@bok3ron554 it is possible to have longer than 8 frames in ComfyUI, using the context option nodes in AnimateDiff-Evolved. Of course this will take longer, for example 16 frames will take more than thrice the time to generate 8 frames, depends on your hardware.

You need to use sampler "dpm++ sde" karras for better skin (as descripted in model's manual) . Today you use cheap version of this sampler - "dpm 3m". And civitai's "dpm 2m" is cheap version too. And use LoRa from model author "Zy Pony Realism Enhancer" it's may help too - https://civitai.com/models/927305/pony-realism-enhancer

@xNzX Thank you so much for answering my doubts. I'll try your recommendations.

(I thought DPM 3, as the name implies, would be a newer improved version.)

@ZyloO can you upscale 2x with the 1x-skin detail? I always thought the 1x meant maximum 1x upscale! :D

meichidarkmixReload - Pony model heavily biased towards Asians by default, but with the right prompts; you should already be able to get good results with Pony Realism. Check other people's pictures (of Asians) for prompt data to help you out.

You can also search for a LoRa of an Asian you generally like the appearance of, then use the corresponding trigger words to get some consistency with that in your image generations.

https://civitai.com/posts/11653363 Feel free to copy my prompt data from this small Asian image sample I quickly put together. "Asian" is broad & slightly vague to determine exactly what you're looking for. Most are vanilla Pony Realism, but I showcased one using Dead or Alive - Lei Fang l PonyXL

@ShadowCell  Thank you very much,

Wtf is 'Google collab'? Provide link for take a look. This model do great images without any 'greenish'.

Are you using a VAE? This model comes with one so don't use an extra VAE if so.

In bottom line of his 'Compedium' tutorial for this model, author gives a link on article, how he make LoRas, based on this model. Take a look - https://civitai.com/articles/6621

use some loras like beauty slider, expressiveH ;)

@Gamalonia ExpressiveH has no description. But how I understand, this LoRa make face emotions (expressions) which often used in Hentai cartoon - clenched teeth, ball-out eyes e.t.c.

@xNzX You can use it for PomyRealism trust me. I often switch anime prompts to realism ;) But yeah your choice ;)

@Gamalonia We are talking about faces quality, as topicstarter asking. "Beauty slider" may be improve quality of faces – but ExpressiveH?

@xNzX Ideed!

Pony V6

@Capitann Does it work in Fooocus 2.5.5?

I am playing around with realism for weeks now locally and I can share what I understood so far.
First of, I personally don't think it is possible to 100% stop the model from generating humans, but you can reduce it very significantly.
The "source_pony" tag in the positive prompt feels like it does already 80% to 90% of the work.
But sometimes you might not want to use it, so you have to think about how key words in the positive or negative prompt will guide the model, to favor specific outcomes.
Below I will explain how I see specific tags.
I wan't to clarify before, that I am not saying don't use them, but rather its important to understand when to avoid certain tags and when its fine to include them, in the positive as well as the negative prompt.

-->a few tags with influence towards generations with humans:
- "female human" / "male human"
- "human"
- "realistic"
- "source_anime" / "source_cartoon" **only slight influence
- "hair" / "mane" **It seems to me that the word mane is linked to the word hair.
- "breasts" / "tits" **If you want anatomically correct mares the tags "teats" or "crotch boobs" should still work in the positive prompt.

It is a balancing act, if you use more tags that can also be used for generating humans, then the chances are higher for generating humans.

-->a few tags with influence towards generations with non human creatures:
- "source_pony" / "source_furry"
- "feral"
- "tail"
- "fur" / "fluffy" / "scales"
- "pony" / "horse"
- "quadruped"
-ect.
The key is balancing.

At last something very interesting I found was that the Photo Real Negative Embedding from "Pony PDXL Negative Embeddings" is heavily trained on reducing ponies. Because when I put that embedding into the positive prompt without anything else I get only ponies xD It might affect the quality somewhat but it is probably very minimal, because I have not noticed anything.

I hope this was helpful. I will keep an eye on your profile of what you will do next, because I have created a collection of most of the realistic generations I could find sofar.

I agree with what DeadHand says good advice. I don't have a problem with this model making humans I would like to add try putting "people" and "human" in your negative prompt. Also make sure your guidance is high enough so the model does what you say. I usually use a 7.5 guidance.

what's your gpu memory? I have a 8GB 4070s，can i use it?

@dbwang931477 yes

It helps a little to separate different objects but works far from perfect:

character description

BREAK

location description

BREAK

style description

It really helps if you have multiple different subject matters, & don't want traits among them to blend together. So for instance; a fat, dark haired man (subject 1), & a thin blonde haired woman, etc (subject 2). A lot of times, if I mention a trait like "athletic" anywhere in the prompt; it'll just spill over into every person in the image generation without the "BREAK".

@MasterArt Ok, thank you!

@ShadowCell Ok it's clear now, thank you! 

The size of the checkpoint file itself (close to 7GB for v2.2 VAE fix here) is a close guess, best to have at least that. Loras will add to it. Clip/Vae usually get offloaded to CPU depending on your software. (keep in mind all need to be loaded into system ram first whether or not it goes to VRAM)

thanks, that's good to know.

@firetome1998 It doesnt need though 8gb Vram. You can use with 4gb too with Forge. It will offload to system ram frequently, but you can use it.

Model size doesnt matter, i used 12gb fp32 models on 8gb Vram with 5 loras(thats a plus too). For loading of models it uses system ram, than it uses around 4-5 gb Vram during generations.

Use this negative prompt:
pale skin. afternoon, night, sunset, fat, cartoon, anime, CGI, 3D model, low-resolution, blurry, noisy, deformed, glitch, overly saturated or desaturated colors, bad anatomy, watermarks, text or signatures, unrealistic proportions, exaggerated expressions, airbrushed details. Avoid any elements that detract from photorealism, such as poor lighting, incorrect shadows, and unnatural skin tones, low-quality, moir pattern, downsampling, aliasing, distorted, blurry, glossy, blur, jpeg artifacts, compression artifacts, poorly drawn, low-resolution, bad, distortion, twisted, excessive, exaggerated pose, exaggerated limbs, grainy, symmetrical, duplicate, error, pattern, beginner, pixelated, fake, hyper, glitch, overexposed, high-contrast, bad-contrast

Some facial expressions showed in this article - https://civitai.com/articles/6621/pony-realism-compendium

@xNzX Yes thanks, I have already check that out. That actually confirms what I wrote :)
Some expressions are rather funny, for example the clenched teeth one ;D
Of course I compare it to the oritinal pony where you basically have every possible expression perfectly.

@ZyloO Yes I know, I did not want to criticise, but only give some productive feedback. The model is awesome and I would not deduct any points because of that.
Ok, I will check it out.

I agree. I just found another one and I am testing to see which one is better. Have you tested this?
https://civitai.com/models/443821?modelVersionId=1478064
Tell me your opinion.

It favours well, perhaps somewhat coming out a tiny bit ahead using just the vanilla model in terms of realism. I think aspiring towards total realism via Pony/SDXL is a mistake though. SDXL/Pony isn't really fully capable of it, & I've used other models that have tried to go that design direction at the expense of everything else. It starts to create uncanny valley like outputs that really cripple it; the more realistic it tries to get, the more & more glaring any mistakes start to appear. & I mean even some pretty subtle mistakes at times, like the perspective of eyes not quite being aligned on the face right, etc. Besides that, I found CyberRealistic to make way more big mistakes, too (multiple hands/arms coming out of nowhere). If you like 'boring' portrait pictures, then it does well though (perhaps better vanilla than PR), but I found it falls apart quickly when you try to do anything really creative with it.

This is why I think Pony Realism is the best. It understands this realism concession well enough; & instead subtely favours style aesthetics that make image outputs look appealingly realistic, but not to such an extreme that if anything appears even slightly weird or deformed, then it just kind of ruins the whole thing & makes it look way weirder than it otherwise would.

PR is also the most versatile - I've never come across another XL/Pony checkpoint that plays as nice with LoRas, ReActor, etc. Pony Realism (PR) is absolutely amazing in this regard! I use facial composites blended together for character consistency across generations (also ReActor face upscaling). I've never seen another checkpoint that just understands this whole concept anywhere near as well as PR. It just 'gets it', & makes the end image generation fit together really well. With other checkpoints; the face/eyes just look as if they've been badly photoshopped on & aren't aligned. Where as with PR, it looks fluid (as if it's part of the original vanilla image generation). I would say this is critical! Use ReActor's face upscale regardless of whether you intend to face swap or not! It's extremely cheap & quick when compared to upscaling the entire image (like an extra second of generation time).

(I was going to go say a lot more, but I'm super tired from being really ill this past week, so I'll wrap it up). Anyway, I have a month old comment on the model page that reviews it (though I've used so many models at this point, that I started keeping notes with each download, just so I'd know all the strengths & weaknesses of each; & how to get the most out of them. I find it hard to not go insane & horde tons of them lol.

So next comment will be a copy & paste of the other comment I posted on CyberRealistic Pony (I'm not sure if you can even link comments). Then I'll post another comment at some other time detailing all the most interesting checkpoints I think are worth mentioning that people are more likely to have missed.

@bobby888 "I've been testing this; switching between v6.5, v7 & v8 back & forth. All the results are surprisingly similar using the same data/seed (this isn't usually the case with other versions I test across for different checkpoints). v8 is coming out slightly ahead as superior though (vanilla) - slightly less busted features, eyes, anatomical compositional goofs (quite subtle to generally detect sometimes, but you just get this uncanny valley feeling with something just being a bit off more in older models, even if it's not outright big goofs with fingers, etc).

v6.5 was my fav of the bunch prior to this one. I feel confident this is the go to model now, enough so to delete my older versions. v6.5 usually made more anatomy errors, but there was just something special about it's aesthetic/style that I didn't want to let go of until now."

https://civitai.com/images/54932204 v6.5 (my fav, but when prompt complexity increases, the checkpoint starts to buckle under the strain more compared to v8 (bad hands/anatomy). Differences tended to dimish more the more complex I made the prompts/more akin to real prompts I'd actually use with all the add-ons).

https://civitai.com/images/54932258 v7 (notice the busted eyes)

https://civitai.com/images/54932358 v8 (the most realistic probably, but I find it just lacks style compared to v6.5 a bit). v8 is way more consistent/reliable as prompt/input complexity increases though, so v6.5 just becomes frustrating to use/fix under real working conditions by comparison.

This is a tiny sample, but I did this many times over. This is a generally good representation of base outputs I found. I typically avoid LoRas for testing, because a lot of times it's not reflective of model strength, but rather idiosyncracies between them just clashing together badly).

Will edit this comment into something better later (too tired right now, but here's a ton of links to check out)

https://civitai.com/models/166609/realismbystableyogi (This is what I was referring to when a checkpoint's reach exceeds its grasp). It looks really good if you're willing to wrestle with it though, otherwise it can be difficult to get really good results with.

https://civitai.com/models/477851/duchaiten-pony-real Still 'realistic', but heavily stylized, & I love the style. It's biased more towards Asians, but this actually works surprisingly well to it's favour when you mix it with caucasian Loras or ReActor composite image blending (where you blend a bunch of different face models together).

https://civitai.com/models/376450/duchaiten-pony-xl-no-score (Variant)

https://civitai.com/models/643227/meichidarkmixreload Similar-ish (also heavily stylized)

https://civitai.com/models/520661?modelVersionId=933040 2DN (I should preface this by saying I don't really like anime typically, but in terms of less realistic checkpoints; I find I often quite like the outputs on this one).

https://civitai.com/models/499918/dr-pibb-ponyxl (Less realistic again, but this one just 'works'. The image outputs just look good without much effort, & it fills in all the blanks in the image prompt really well, & without tons of weird stuff going on in the image).

https://civitai.com/models/153568/real-dream (Similar to Pony Realism, but not as good).

https://civitai.com/models/194768/jib-mix-realistic-xl Jib Mix Realistic XL (I haven't had the chance to use this one much, but I'm liking what I'm seeing so far)

https://civitai.com/models/277058/epicrealism-xl epiCRealism XL

https://civitai.com/models/488361/speciosa-realistica Speciosa Realistica

Sadly there's a bunch of other models in my permanent collection I'd link, but the creator's removed them ;(

Honourable mentions for FLUX

https://civitai.com/models/686814?modelVersionId=1126217 Jib Mix Flux

https://civitai.com/models/778691?modelVersionId=1129380 Fluxmania

https://civitai.com/models/118111?modelVersionId=982160 Copax TimeLess

https://civitai.com/models/938037/copax-flux-realistic Copax flux realistic

https://civitai.com/models/978314/ultrareal-fine-tune UltraReal Fine-Tune

@ShadowCell Good job with all the tests. By now I have also tested it quite a bit and it became my favorite model. I have the best results and control. I guess I was lucky finding a great setup.
here is an example image https://civitai.com/posts/14042302
Locally though the face is better for some reason, but a really great model.

@bobby888  Wanted to check it out yesterday, but civitai was a huge lagfest for me...even deleted some of my work/edits ;( I'm guessing the new "no model left behind" transition was to blame for network traffic collapsing.

Anyway, it's nice to see CyberRealistic Pony has a new model out (already downloaded, just need to test it out now). Your linked image looks good btw :) It looks a little blurry, & the eyes look a bit hazy though.

https://civitai.com/models/118427/perfect-eyes-xl Try this for eyes (sometimes better to not use though, but can be lazy fix in some situations)

https://civitai.com/models/341353/expressiveh-hentai-lora-style This if you get tired of the same blank facial expressions

https://civitai.com/articles/12420 Try this to remove blur, or pass to extras tab & use Latent or Lanczos upscale (both virtually free options).

Using "Closed Mouth" in the positive prompt can also fix a lot of headaches needing to use inpainting, etc, when you get weird teeth, also.

I'm intending to write more advanced guides to increase image quality cheaply (or not cheap via FLUX or intensive upscaling options).

@ShadowCell I created that image in Civitai. On my local setup I haf adetailer and also use those loras. Only the Expression lora doesn't work that well in pony realism, and I just removed it, but that's fine for me.

which version is the best in your opinion?

@challasvarias01106 For Pony Realism I use v2.2 Main + VAE, & v2.1 (start with v2.2 if you only want to download one). As for CyberRealistic, try the latest v8.5 (I'm liking it, & have tried most of the prior models).

https://civitai.com/articles/12416/resources-worth-checking-out Here's a rather long list of resources I've compiled & used (with some personal rankings). Naturally Pony Realism is at the top of the SDXL/Pony list ;)

Edit - Oh, I should point out that Pony Realism/SDXL in general generates images for me in like 10s typically (so I don't even consider turbo variants for SDXL/Pony), but if you're lacking VRAM/weaker PC, then the other options might be more suited for you. Also, when I used AUTOMATIC1111 (CUDA 12.1, I think) my generation times were more than double btw! Forge is a fork of it, & nearly identical (but still getting updates). It's less compatible, but newer CUDA versions mean faster generations.

In Forge; you can tick the ReActor checkbox to enable it, then the Upscale tab there in ReActor, then tick "Force Upscale". This adds like a second or two to the end generation time, but will fix blurry faces, & massively improve the quality at virtually no computational performance hit. You may need to have an active image still there for the 'swap' you don't actually intend to do.

Remember, in Forge/AUTOMATIC1111, the 1st detected image from the left in the "Target Image (result):" is 0. So in other words, 0 = 1st image from the left, 1 = 2nd image from the left that will get swapped, 2 = 3rd image from the left of the image you're generating, etc. You can also set upscaler option to "Lanczos" (This is very cheap/quick), though you can wait longer for better results using something else, too.

When you deliberately don't want to use an actual face swap (but just upscale face) just set "Target Image" to 9, or some other number that won't get used.

@ZyloO The results are less erratic, & much easier to predict/control. Whole thing is automated, & takes 2 seconds to enable also. It's more suited to someone just starting out where you just change one setting & instantly improve image quality with no hassle.

Unfortunately, I don't think you'll be able to see the comparison gallery yet to compare lol (seemingly taking forever to unrestrict the first two images...even though they're virtually the exact same). There's other options to improve results (better even), but this one doesn't really have any performance or time trade-offs. The base image quality resolution is also really low (intended for entry level users/hardware).

I'm surprised to hear you've never used ReActor. Back then I just downloaded whatever...a lot of it trash lmao. I got some super bad advice starting out (especially a lot of "this checkpoint is great" suggestions = sd 1.5 junk from YT vids published like 1-2 months before FLUX came out lol). By far my biggest mistake was wasting time with SD 1.5 right off the bat. What a bad experience....It was so damn fickle & prone to 8 armed monstrosities & other weirdness lol. Was Kind of a trial by fire though, I guess ^^

(Also the cult of Automatic1111; by that point it was just dated & obsolete rly. People shouldn't have been suggesting it in new vids at that point, especially with things like Forge more or less virtually identical, but with massive performance improvements, etc). Also Comfyi was bad advice for absolute beginners, too.

@ZyloO Damn, this is all so complicated for me, maybe there is some guide or you can tell me, I downloaded your model, 6 gigabytes, downloaded ComfyUi, then because of ignorance I downloaded stable 1.5, and in the extra_model_paths file I made a link to the 1.5 model, the results are so-so, probably because of this? How can I just run your model 🥲

@ZyloO I have used other Ai that work through the site, I just need to understand how to install and where to download, I have already downloaded your model

@ZyloO Thank you, and sorry for disturbing you, I'll try now, the last question, in ComfyUi, I open the checkpoint of SDXL, right? And I write PonyRealism at the beginning with the prompt: <lora:ponyRealism_V22MainVAE:1.2>. If so, then the picture quality is good, but not at all what I need, or should I just choose Lora Pony in the checkpoint and that's it?

@ZyloO Yes, I figured it out, thanks for the quick answers. The result was awesome ❤️

@jakijons Friend, I just recommend you to watch the videos on YouTube on a1111, and here in a nutshell, very few people can explain to you. Everything is clear in the videos, watch the videos!

@AlexVetrov I've already figured it all out, thank you.

SwarmUI is another option btw (just chucking it out there for other people reading), because it has a fully working ComfyUi backend that you can switch over to in 2 seconds, but it also has a wrapper that's more like AUTOMATIC1111/FORGE though, so it's perfect if you want the best of both worlds, & it gets more frequent updates/support than Forge.

Stability Matrix is another option where you can try everything out (but I haven't really bothered with it myself).

yeah i noticed the same thing, too bad :/

@ZyloO what is that

@ZyloO thx

@ZyloO thank you

I don't really find illustrious or noobai better with hands tbh, they do seem better at sticking to the prompt, but this is still the best realistic model I've found by far. Still using it a a lot. Even if I used other models for prompt I'll hires and adetail with this

@ZyloO I apologize, it wasn't the model, it was whatever Civit did in their update that they just reverted (I didn't get no announcement notif of the update, and the only resource that had updated was the model, so I assumed it was the model), again, that's my bad, but hey at least I can publicly admit when I'm wrong

@ZyloO idk if my reply sent or not, but check the edit

@ZyloO Do I understand correctly that after I download it, I will need to add ComfyUi nodes: 1. Load upscale model. 2. Image upscale. 3. Sampler. 4. Image?

Here´s a Link to an youtube-Vid, showing how to use Upscaler in ComfyUI:

https://www.youtube.com/watch?v=AW6BdxyqK8g

@Draftsman appreciate it

Here is a Link to the "AI Playground User Guide":

https://github.com/intel/ai-playground/blob/main/AI%20Playground%20Users%20Guide.pdf

...and here is a Link to an Youtube-Vid, showing how to use Checkpoints in "AI Playground":

https://www.youtube.com/watch?v=1FXrk9Xcx2g

tyvm , you are amazing

I guess, with "background blur effect", you mean "Depth of Field" ? I remember, that there are Lora´s for that, here on CIVITAI. Search for it. Good Luck ;-)

@Draftsman hanks, I'll try.

put "bokeh" in your negative prompt.

@wild_RPG thanks for the reply, I tried, but it didn't help.

blurry, motion blur, blurry background, depth of field in negative

@donaldvanderlinden121 thanks for the reply, I'll try!)

@ZyloO I have send you a pm.

You need to fine-tune (post-edit) your images in Photoshop -- substitute colors. Beauty requires sacrifice )

I never tried 2.1, why you think its better?

@Draxx318is I like it better.

I remember having used 2.1 quite a bit in the past. But 2.2 always has issues that require extra prompting or Lora. Too expensive on this site. Looking at the examples I see a lot more issues than in other models.This version however is very good with light control and also works well with character Lora (a lot of models I tried try to force blend their own default face with a character Lora resulting in wrong faces). Interestingly three of the asian models, especially WAI_xyz works very good.

Denoise strenght of ADetailer == 0.5 (near the same face).