I recommend DPM++ SDE Karras, CFG 5, 30 steps.
If you are looking for absolutely cranked skin textures, I recommend v2-rc over v3. But v3 offers better consistency for hands and eyes.
I recommend trying out portaito as well.
Description
This model version achieves higher detail; especially in regards to skin texture.
But it at the moment does so at the cost of some stability and prompt adherence. So for now this is only a release candidate that you can try - if you find the model hard to work with use v1 instead.
FAQ
Comments (46)
lookin beautiful bro! keep it up
Thank you so much! Any feedback good and bad is very welcome to help improve further
Wow, you definitely have a good eye for model mixing and it shows what your aesthetic aims at. Great models, new one is above many others. Hats off to you!
Definitely agree! The realistic visuals of this checkpoint are top notch!!
Thank you! Any feedback on what weaknesses and strengths are helps me continue to improve, all feedback welcome
awsome! best model I tested for my case
Thank you very much! Any feedback as to how it helps your use case would be super useful to continue to improve my approach
I just ran a series of various tests to measure 13 models against each other. The test used a scoring system of 1-5 (1 being the lowest). I tested prompt adherence, accuracy of the picture itself, ability to work as a refiner, detail and complexity of scenes, quality of final render, and several other little things.
This model scores second place, and truly caught me by surprise. The model scored second highest in scenery complexity, very high marks for realism, pretty decent prompt adherence, excellent and somewhat randomized face generation, and overall creating beautiful eye pleasing images.
Of note, this model is excellent with a refiner, utilizing both real and cartoon models, due to the complexity of the scenery, poses, and general quality of every image. Further, this model can be used itself as a refiner on the backend.
Truly an underrated model, that I believe conclusively is better than many others out there. Keep up the excellent work.
Wow! I can't express enough how motivating this is to hear! Thank you so much for taking the time to write this, it's very appreciated.
I'm glad to hear it seems to work well, and it further motivates me to continue to develop my custom model merging algorithm.
There are many stones still left unturned and I feel like I have just gotten started; I'm very welcoming to all feedback and information as this helps me guide how I invest time into problem solving. So far it's been really fun to publish the work in progress for my model merging techniques - but yet to come is work I've done specifically around text encoder training - which is something I believe is holding back the Pony/SDXL space at large.
I'm very interested to hear that you've run these meticulous tests - please don't hesitate to reach out to me on chat!
@epiTune my friend, I think it should be noted that prior to me testing these 13 models, I had probably sifted through about 50 or more before I got to this point.
On another note, something I forgot to mention, it scored SECOND on randomly generating faces, only being beat out by a few decimal points. Of all the models I tried, nearly every single one would give 1girl pictures that all looked exactly the same, with minor variations of hair color. This model does NOT do that, and makes it an excellent refiner for things like faces, bodies, etc.
My method was devilishly simple... I would just type, "Portrait, Female" into the prompt, and this was able to generate different faces every time for the most part. I can't stress how much I get tired of seeing the same faces. This rang true for selecting different hair colors, etc.
This model really stands out to me, and caught me completely by surprise today.
@RedBjorn This is so fun to hear! Is there anyway you could share your test results with me? I believe it would be invaluable and would help me improve a lot!
I agree this is an amazing model, I wrote to epiTune right away telling him how amazing it is, not overly saturated or overly exposed. Really amazing model. @RedBjorn , what is your #1 model that you are talking about?
@porsche9111 the model that scored highest is:
https://civitai.com/models/592380?modelVersionId=842050
Specifically the lightning versions. It makes an excellent base for refining, or quick image building. It misses out severely in random face generation, and other areas. Combining these two models, using one as a base and the other as a refiner produces some amazing results.
As with anything, there are pros and cons to each model. Ultimately, I wanted to find out what the strengths of each model was.
Post the test results. Would be cool to see them.
@StanleyPain I've considered it, but I already wiped it from my hard drive to avoid the temptation of sharing that information. I'm trying to avoid the black hole of developing a deep regimented system, and further, have it scrutinized by the public. I simply don't have time. That being said, I may in the future do another few tests... But, the result of the testing has been extremely positive for my needs.
This is great work. Very flexible, accurate anatomy, good prompt adherence, and a great style. well done!
Thank you! Any feedback helps improve it further
I wrote-up a review, but it seems it's been lost on adding images... Anyway, v2rc is savage. Especially if you don't use score tags. It can be ugly, but for a Pony checkpoint it's a virtue. It comes with some amount of chaos and some AI noise, but got a character. To show it, omit any "quality tags" and don't use LORAs. A wild ride.
The gens you've posted in the gallery truly have character and a unique bend to them; really fun and interesting to see!
None of your samples have any metadata? Are we supposed to take this model seriously?
If that's a deal breaker for you then you're of course fully entitled to look elsewhere; there's no expectation that you must take this model seriously - it's just a fun experiment I'm sharing as part of things I'm working on. There's images in the gallery which have metadata. But yes I will start uploading metadata for the images I generate; as multiple people have requested it.
@epiTune I know it stinks when you post something and people complain. And I hate to be a complainer.
But honestly, it kinda looks like you're hiding something or don't want us to be successful, or something. It's extra steps to remove generation data. And 99% of the models here include metadata.
It's helpful to have to know how to use the model. Do we need trigger words? Do we need score tokens? What does it take to do the very basics? But we can't even see if you're using loras to get these baseline results.
Regardless, thank you for sharing. I will give this model a try.
@thenewgray Hello, appreciate the thought through reply.
I must clear up that I don't go through any process that intentionally removes metadata.
I do all my generation using the diffusers library https://github.com/huggingface/diffusers with code I have written - and I simply have been lazy and haven't yet implemented adding metadata to the exif parameters (which is what tools like A1111, Comfy, InvokeAI, etc - all do). So yeah, shame on me for not doing that yet. But I have never scrubbed metadata from any images I have uploaded - it was never added in the first place; that's the issue.
It should behave like most other Pony models and thus benefits strongly from score tags, yes.
All the images I have uploaded are "vanilla" - no loras. The images come straight out of a basic SDXL diffusers pipeline using this:
output = pipe( prompt=prompt, negative_prompt=negative_prompt, width=var("width", 832), height=var("height", 1216), generator=g if seed != -1 else None, num_inference_steps=var("steps", 30), guidance_scale=var("cfg", 5.0), callback_on_step_end=sdxl_progress_callback() )
So couldn't be more vanilla that that.
For good results I recommend pony style prompts (this includes score tags in positive and negative)
DPM++ SDE Karras
30 steps
5 cfg
@epiTune This is useful information. Thanks for taking time to reply.
Very good model, over the top!
One of my favorite real pony models by far, thank you!
So nice to hear! What do you like the most about it?
@epiTune The prompt adherence and creativity perfectly balanced with very good realism and backgrounds, so many models can feel like a bit of a battle trying to get your vision across to the model, but Opticon is always a joy to prompt for, its also perfect to use in merges, especially if I find Ive overbaked a merge and the prompting is getting a little broken, I just add in a dash of opticon and were back on track, this model definitely inspires me Epi, incredible work! Im currently seeing if I can improve upon the things your model does so well even more with a merge Im working on now
Could you please add at least one prompt of one of your images examples
score_9, score_8_up, score_7_up, 1girl, solo, <whatever else you want>
score_6, score_5, score_4, censored, source_furry, source_cartoon, blurry, lowres, watermark
DPM++ SDE Karras
CFG 5.0
30 Steps
There's nothing special to it.
@epiTune thank youuu
@epiTune Thanks epiTune. It isn't about being special. It is about having a verifiable baseline. A LOT of models have demo images that can not be reproduced. It's misleading and wastes people's time.
@Ruoar All showcase images for v3 has metadata : )
This is an incredible model!
Amazing response to lighting prompts!
Great realistic body anatomy.
Great skin rendering, very realistic, doesn't the the issue of looking like plastic.
This is becoming my new favorite realistic model.
Super fun to hear! If you have feedback on anything that could be improved please let me know!
And this is by bar the most realistic/best fabric I've seen any realistic model output so far, especially something like semi-transparent silk type of material.
One thing I noticed is that this model tends to stay on the darker, lower exposure side, which can be great for achieving certain realism. But sometimes I want the scene to be brighter, say if the character is wearing darker clothing, I find it hard to get it to lighten up the scene. I tried different sampler, CFG, Lora and lighting/environment prompts.
if only civitai could fit this on site gen. textures look great
What an incredible model! I'm impressed with your work. Thank you!
P/S: Can I fine-tune my dataset with your checkpoint using koyahSS and Dreambooth?
Yes, there's nothing preventing you from doing that 👍
thank you for this one, keep going
When will you update this model? I can't wait to see it. :D
I should get back to it, for sure. But my main focus has recently been Illustrious - which straight up is a far better model than Pony.
@AntiLeecher v3 is out now : D
@epiTune Thanks. I saw it and made a tip already. :)
This is still a damn good model.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.