Join The Tinkerer on Whop to download this model and many more — all for one monthly payment. Plus, get early releases, private tools and a lot more. 👉 Join on Whop
CyberRealistic Pony blends all the charm of Pony Diffusion with the striking realism of CyberRealistic. The vibe? You get everything from adorable to bold (sometimes both at once) with crazy-detailed textures, moody cinematic lighting, and a hint of AI flair.
🧠 How to Use It
Sampling method: DPM++ SDE Karras / DPM++ 2M Karras / Euler a
Sampling steps: 30+ Steps
Resolution: 896x1152 / 832x1216
CFG: 5
Clip Skip: 2✅ Recommended Prompts
score_9, score_8_up, score_7_up, (SUBJECT), Negative Prompt examples:
score_6, score_5, score_4, (worst quality:1.2), (low quality:1.2), (normal quality:1.2), lowres, bad anatomy, bad hands, signature, watermarks, ugly, imperfect eyes, skewed eyes, unnatural face, unnatural body, error, extra limb, missing limbsit's also possible to remove the normal pony tags:
(SUBJECT), Negative Prompt examples:
(worst quality:1.2), (low quality:1.2), (normal quality:1.2), lowres, bad anatomy, bad hands, signature, watermarks, ugly, imperfect eyes, skewed eyes, unnatural face, unnatural body, error, extra limb, missing limbs🛠 ADetailer Settings
Adetailer model: face_yolov9c.pt
If you only want the main face being refined set 'Mask only the top k largest' to 1.☕ Support
Enjoying the ride? [Buy me a coffee] – but only if the Pony delivered something you actually like.
⚠️ Disclaimer
This model might generate sensitive content. Whatever you make with it is on you. Don’t do anything weird and try to blame the horse.
🔗 Links
Backup location: huggingface
💡 Need Better Prompts?
This custom ChatGPT was made to whip up top-tier prompts just for this model:
🔗 [Try it now on ChatGPT]
Description
The Evolution Continues
V14.1 is here, and it's a direct upgrade from V14.0, shaped by your feedback. This version tightens up how well prompts are followed, smooths out the style, and brings better consistency overall. You’ll get results that feel more natural and in sync with what you asked for, no matter the vibe.
What's new:
Improved prompt adherence and tag consistency
General polish and optimization based on community feedback
FAQ
Comments (63)
Early Access really is ruining Civit.
Only for people too impatient to wait 15 days and who don't understand that developing checkpoints costs money and time.
They have to keep the lights on somehow... you think hosting is free?
@SubtleShader Well, paywalling content is never the way to go when you're already using shady assets to begin with. I didn't mind contributing to authors sending some buzz their way, but I will never donate/contribute to someone locking their content.
Has nothing to do with impatience.
@phlesh I know very well hosting is not free... Locking content behind a paywall is cringe and probably does more harm than good.
@Madsen1981 There is no paywall option on Civitai. In case you did not notice. The only way is to not upload here.
Run it locally!
@Madsen1981 Just to clarify - all my models eventually become public here on Civitai, so nothing is ever permanently locked. If you don’t want to wait, you can join my Patreon anytime for full access to everything, or pay a small one-time fee here on Civitai. Patreon members simply get early access, plus many other models, experimental builds, and extras as a thank-you for supporting ongoing development.
Creating and maintaining these models takes a lot of time, testing, and resources. That Patreon support is what helps me keep improving and releasing updates like this one.
If you prefer to wait, no worries - the public versions always follow later.
LOL you can use the older version until the new one is publicly available or simply just pay, i honestly don't see anything wrong with it. I'd do the same if i was creating checkpoints and i am simply waiting for release of 5000 super series, i'll be creating checkpoints too.
lots of people have a life outside of waiting for AI models to release. just be patient geez
How so? Still going to come out to use,
@Madsen1981 That isn't even what the word "cringe" means... there is nothing embarrassing or "icky" about it. The only thing cringe is when people whine over a 2 week waiting window that stands between them and trying to level up their AI porn habit. Now THAT is hard to stomach.
@phlesh lol
what realism and skin lora do you guys recommend? and weights
I use Pony Realism Slider by Klaabu with a weight of 2. Usually, that's enough to get decent-looking skin
V.11 was the last one version who can produce accurate face details for face LoRa and as accurate for making futanari without any problems or Futanari LoRa .. i struggled with all this starting from V.12 to the latest 14.1 unfortunately the look realistic but never accurate on what i mentioned🥲
Самая популярная модель, похоже
As others have mentioned, version 11 is the last version with good prompt adherence.
11 was the best version
Wow, just tested every version from 7 and above. But you're (OP) spot on! I also have the feeling that the colors are less vibrant and there is too much contrast in v12 and above.
In Cyberdelia's defence I do have to add that I only tested with LoRa's. Later versions seem to handle those differently. Lowering CFG on LoRa's might help slightly.
From what I have tested so far, v 14.1 is a masterpiece.
wait I just noticed that you have a full model for purchase and Pruned Model, what's the difference? for now going for the pruned as it's what I saw everything else has it
For general use it doesnt matter. fp32 full is more like for people that want to remix checkpoint, merge stuff in and such..
@Mescalamba I see, thanks!
@AinzAkaOverlord You can still use it, in case of ComfyUI, it will automatically become fp16 when loaded. Unless you have something with like 16GB VRAM, then you can enable using fp32 models, but since ComfyUI I think simply doesnt work in fp32, even when persuaded really hard, it wont have much benefit.
Tho in general, CLIPs in fp32 are a bit more accurate, tiny bit extra detail, less errors. If you must or want to have that extra 5%, then sure.
That you can do without loading whole fp32 model, there is starting argument for that too.
can I run it on Fooocus Collab? I`m a beginner
Is there a difference in quality between the FP16 and FP 32 versions?
FP32 is nice if you want to use it for mixing with other checkpoints etc. For normal use - use FP16
@Cyberdelia Thank you for your answer, but are the renderings better if I use it alone or with loras?
@Suzanne That's a different question. I almost never use Lora's. The model can basically do a lot. Sometimes it can make the result worse if you use LoRas.
@Cyberdelia Just a suggestion, but think it would be a good idea to check out LoRA adherence when creating a checkpoint. Noticed that some of my character LoRAs don't work anywhere near as well with v14 and v14.1 when they were working great with v13.0 (which imo had some of the best LoRA adherence out of all the CR Pony and Illu versions to date).
But other than for testing, I do appreciate when a creator doesn't use LoRAs (or embeddings) when they post their samples. It's more impressive when a model can produce amazing results on it its own. Anyway, keep up the great work!
@AFD_0 I honestly don’t see why versions 14 and 14.1 would have worse LoRA compatibility than version 13. Normally, LoRAs can cause issues if they’re merged directly into a checkpoint, but that’s not the case here. I’ve also done some testing myself and didn’t encounter the same issues you mentioned, so I’ll need to look into it further this week.
Regarding the sample images without LoRAs - yes, that is intentional. You can’t really show the true quality of a checkpoint if you’re using LoRAs to enhance it. I should also mention that I personally rarely use them; I think it’s important that the base functionality works well on its own.
@Cyberdelia I'm still messing around with boosting CFG and maybe even pushing LoRAs above 1.0 strength, but they worked much more effortlessly with v13.0 for some reason. With v14/14.1 it feels like a slight battle between getting better realism with lower CFG and better prompt/LoRA adherence with higher CFG. It's weird how some models do character LoRAs really well while others will excel at style or scenario LoRAs. No idea why that is. Anyway, was just something I noticed and they're still both very excellent models on their own without any LoRAs.
@Suzanne I've been using @Stable_Yogi 's Realism LoRA v2.0 Pony version with CyberRealistic (and most other models) at a low strength of 0.2 to 0.3 with good results. Best thing to do is try one (or two) out with different strengths and see what you get. Sometimes a creator will build their checkpoint with the same or a similar LoRA already merged in and adding the same back on top can often produce some really nasty and unpredictable results.
@Cyberdelia using FP32 makes it better at upscaling itself without any visual artifacts, from my short experience.
@mednorator thanks
@mednorator Ok, I only use FP32 - but I will check if upscaling is better with FP32 then FP16.
Do you do seaart?
I'm actually getting much better detail with v14.0 than v14.1 when upscaling.
Any plan for cyberrealistic chroma or qwen?👍🏻
Chroma would be friggen awesome.
Very cool
Soy nuevo y no se cómo funciona esto ? El que archivo que pesa tanto que es ? Una IA?
Busca "stable diffusion civitai como usar" en youtube.
Love the model, I find I need more and more to use face detailer, in swarmUI I add :
<segment:yolo-face_yolov8m-seg_60.pt-0,0.20,0.25>
And find it helpful while avoiding creepy plastic doll territory...
One question, it seems eyebrows are not matching hair anymore?
Great, but very frustrating in use. My main problem is prompt adherence. Some words have strong weights that break the image and higher resolutions very quickly introduce abnormalities. CFG is very sensitive, leaving no room to play. Overall, you are able to get really beautiful images, but I guess you must use Loras and upscalers to have sustainable results. Out of the box experience is quite tedious. I found other models easier to work with.
I totally agree. It's nice to hear someone else say it. Which models do you prefer, considering this?
Tried this for 5+ hours at this point. Great results with minor prompt tinkering and no loras. I was happy with these settings (comfyui):
steps: 28
cfg: 3.6
dpm++ sde karras
Couldn't get every pose I wanted, but that was expected. Great job!
Can you share you coumfyui workflow please
I feel so noob. I work with comfyui and wan22. There are many workflows, BUT, I can’t find a workflow for this pony model… can anyone share link to a simple workflow for this?…
Same, don't know how to utilize this
I don't think it necessarily requires a complicated workflow. Even Comfy standard template works great.
it takes a lot of downloads. at least 67
@freestuffpl0x42069 ?
@SoloTouching Have you tried turning it off then on again?
This is a fantastically realistic model! I use it quite often and am always surprised at how well it conveys realism. Bravo!
你能微调一个亚洲版本的模型吗？
默认生成的亚洲人都很丑😂
can you make a new FP32 version , the FP32 of V13 from HF was very good
New FP32 of what?
V15 is the ultimate Pony checkpoint for realism. 💯
V15 mixes in best of classic Pony style and the experimental laboratory of Catalyst path slowly refined to maximum potential, this is the laboratory and you are the scientist. But you actually get amazing results! Photorealism can go as high as your prompt & workflow settings ceiling allows it to go. Can render easy & fast stunning images upscaled to 2K or 4K.
Quality surprised me, previous ones that i tried was just couple days ago, the 14 and 14.1 so i already had nice settings and proper workflow for this already in my Comfiest of ComfyUI. with V15 can make amazing images, photorealism is at level, when does images with backlit sun, the sun shinethrough effect from a woodland dryads flower crown with leaves did shine through and show the midrib and veins of the leaf perfectly as in real world photography. The detail level is very accurate, same for skin and eye details when you have your settings properly done. Very good eye catch lights/other reflections and also iris pattern detail. Can get very good subsurface scattering effects visible.
Yes, you can make VERY photorealistic images with this, that does not look like: "Hi, my name is 1girl, i am the Ponygirl with the Ponyface!" but actual phone camera-like images from real world, environments can look very lifelike that does not look like "AI slop", but actual real environments with variety and not like "tiles" (you know, same grass simplified backgrounds etc):
Need your Lora stack as always with Pony models.
All my usual favorite Loras work well it it, both DMD2 based workflow with lcm+exponential typical settings work 100% and so does the recommended normal non-DMD2 settings, of course once you find the perfect settings for DMD2 that has the same quality and variety as non-DMD2 workflow, then it's often hard to go back anymore.
Is this (as any Pony tbh) easiest for someone new trying as their first checkpoint at home? No. Only because Pony in general for best results demand that stack of well selected Loras. I currently have 11 Loras in my Powerloraloader node always when i start making images in Pony, and that is the start, so can get to more than that. :D But results then are same photorealism as some SDXL checkpoints but with the control of Pony.
This works well with tags BUT, i tested some cool normal language prompts, and those worked nicely also. So does mix and match, so can put the quality modifier spam in it if wishes after a description of a situation you want to see in normal language (ask AI to make it in optimized format if you are still new, you'll then see and learn fast what works so can then just use modules of that on your own). You want to see pretty landscapes, right? Right!? You naughty boys, i know what you are rendering.. :P Oh but it does well in that, i tested, guys it was for science i promise uwu.
What i use: Almost standard ComfyUI workflow +
PowerLoraLoader
ClipTextEncode with BREAK syntax
ClipSetLastLayer (obviously)
FaceDetailer
HandDetailer (sometimes)
UltralyticsDetectorProvider (for the face_yolov8m)
Normalized Attention Guidance
Load Upscaler model (for that Remacri 4x or 2x a bit more light and even faster upscaler)
Upscale Image (using model) that makes the previous one do the magic obviously.
There, quite simple. Okay, maybe it was not exactly standard ComyfUI, but close enough. 😅
Amazing work, love especially how well it does with many surface effects, like water ripples, lightings and scattering. Really good with high dynamic ranges, godrays and particles/hazes. It's like 14.1 mixed with Catalyst.. well duh, since it is. ^^
Can also make cool translucent or partial objects on subjects. Yes guys, that means you can make for example wet clothes or slightly see through a dress in the sunlight when the sunshine emphasizes your subjects body contours from under the clothes on the beach on that sunny summer day in natural photorealistic way. That is what you want, that is what this can do with the right settings and prompts. ^^
Big thanks to @Cyberdelia for your experimental hard work to improve things version at a time, love and care for the game shows here in these checkpoints. 💜
V15 is absolutely amazing!
I am getting absolutely stunning results with this new version.
Great work!
how is prompt accuracy/adherence?