Dunno, but it seems as permafixed ponyface at least. Will know more when its available for testing, maybe prompt follow is better and such.

@Cyberdelia right, but... what kind of soul? a darker soul, a happier soul? soulful? soulless? we must know!

Ok, so this is a bit of a laugh......but the same prompt as all the other images literally comes out as a black & white garbage image with v18 lol. I guess this is more of a radical departure than initially thought. Evidently something in the negative prompt field is causing the issue (I've already tried deleting it, but it kind of defeats the purpose of comparison when I do....).

I will have to find the exact word(s) causing the issue & come back to this lol (I was just going to quickly do these images b4 sleeping, but this will take a lil' longer than anticipated).

It's still a somewhat valid comparison still if I just remove one problematic word, but keep everything else the same.

make-up, bad hands, bad anatomy, ugly, deformed, (face asymmetry, eyes asymmetry, deformed eyes, deformed mouth, open mouth), long neck, freckles, moles, elongated abdomen, muscular, muscles, beard, six pack, obese girl, deformed, deformed face, low quality, bad quality, worst quality, (drawn, furry, illustration, cartoon, anime, comic:1.5), 3d, cgi, extra fingers, (source_anime, source_cartoon, source_furry, source_western, source_comic, source_pony), deformed teeth

Full negative prompt is above; "(face asymmetry, eyes asymmetry, deformed eyes, deformed mouth, open mouth)" When I delete this it fixes, but when I delete everything but this....it also fixes it lol. Evidently it's part combination of words added together.

"face asymmetry, eyes asymmetry,"

These are the offending words in the negative prompt that brick the image generation (namely "asymmetry" itself).

Perhaps Cyberdelia can offer more explanation. I will continue with comparison with just those words removed; everything else the same.

v18 seems very vulnerable to some words in the negative prompt field; another image generated with weird distortions using slightly different negative words.

So v18 comparison images aren't the best of the lot.... I feel bad even uploading them for numerous reasons.

1. Because I don't feel it's an entirely fair comparison (people may draw unfair conclusions based on so little).

2. Because it sends the wrong message; namely, don't deviate from the formula, because it will very likely result in bad image outputs. The original comparison images were tailored specifically for those prompts (I can't stress that enough)! When I compare them to anything that's radically different, it's extremely unlikely they'll come up superior as a result of that.

3. Before anyone draws unwarranted conclusions, I think the fair thing to do is to play the strengths & advantages of what v18 offers; not by comparison to years old checkpoints. It sends the wrong message otherwise; namely that trying anything new/different = bad (which is not what I think anyone wants).

4. I'm super tired right now lol, so I'm going to come back to this tomorrow, & will expiriment with trying to produce the best images I can with v18 then. I think this is a better 'test' to see if it's actually an improvement over past checkpoints or not.

Thanks for taking the time to do this and share

So honestly, the comparison images suck a bit for v18 (not really hugely surprising if its moved away from older/previous checkpoints more dramatically).

But I'm liking the bonus images/changing the prompt up more radically for v18. So far I'd say v18 is a welcome mix-up & pretty good so far ^^

It is capable of generating really quite realistic faces & skin, for instance. I also like the tendency of v18 to generate more vibrant colors & sharper images.

Thank you for your work

@ShadowCell >>Before anyone draws unwarranted conclusions, I think the fair thing to do is to play the strengths & advantages of what v18 offers; not by comparison to years old checkpoints. It sends the wrong message otherwise; namely that trying anything new/different = bad

That's fair, but removing/diminishing the behavior, quality and expected results that many have grown to love with CyberRealistic Pony should still equate to an appropriate opinion (or valid comparison) to prior works. I'm liking v18 a lot more than v17, but so far less than some of the prior iterations (v16, for a more recent example). For the most part, it's super-easy to tailor a prompt to achieve amazing results in v18, but it's beyond frustrating when a character LoRA (or outright description) bears little to no resemblance or you use an incredibly concise and to the point prompt and it continuously ignores something important/weighted while exaggerating something intended as a subtle detail. Making an amazing image is always nice, but ultimately meaningless if you can never make the image you are wanting to see.

Maybe I'm just outgrowing Pony, now that modern XL models are becoming just as capable of doing similar things. I absolutely love @Cyberdelia 's CyberRealistic Desire XL v3.0, which is truly a masterpiece among XL-based models. It simply does exactly what I ask of it (and maybe more importantly, what I ask it not to if needed), follows most LoRAs as intended (whether character likeness or anything else it was intended to accomplish) and easily churns out amazing results regardless of how long or short (or batshit crazy) your prompt is or how many random LoRAs you want to throw at it.

And there's nothing at all wrong with a model taking a new direction or being experimental, but I'm guessing it's probably difficult to surpass yourself when you've already set the bar so incredibly high with so many of your previous models. And my goals for what I want out of a model may not always align with the creator's goals for what they want their model to be, and that's.. okay. :)

Tldr; I'd really like to see a future CR Pony v18.5 that has all the gorgeous image quality and creativity of v18 CoreShift, but with the realism and amazingly well-behaved prompt/LoRA adherence of CR Desire XL v3.0

Looking back again, I actually quite like v18 https://civitai.red/images/130161453

I think I might of been a bit harsh on first impressions (perhaps being tired as hell, & half falling asleep at the keyboard trying to get it done before going to bed impaired my first impressions of the images somewhat lol).

I think the Bonus Images def turned out good though <3

https://civitai.red/posts/28479847 & https://civitai.red/posts/28480346 I especially like.

It's funny, too; because sometimes my favourites of the original comparison images change a little, too.

It'd be cool if everyone rated at least 1 or 2 of their favourite images of each; just so we could see which checkpoint/image style the majority like the most.

i have to say that v16 is my choice

@ShadowCell am I right, that all gen faces will be some sort of ideallistic fashion types, and most of pretty other types, that dont match to teoretically ideals, wouldnt appear at all? And is there any tips of prompting more variable or strict faces gen? THX for great work bytheway)

@AnonNameMetoo "ideallistic fashion types" Given image generation checkpoints require literally billions of images to train (exact numbers for SDXL base were never given).... you can see the issue. They need high quality images to train (to get good image generation results back), so large parts of the training data is going to be professional fashion photography, etc, that of course biases it massively.

That said, it requires some aggressive prompting....lol (but can be done).

Try

"Weathered, ugly, fat, old, poor, asymmetrical face, skin blemishes, uneven skin tone, minor wrinkles, imperfect teeth, un-styled hair, average-looking person, middle-aged commuter"

"amateur snapshot, phone camera photo, captured on an iPhone 8, disposable camera shot, candid paparazzi photo"

In your negative prompt: supermodel, flawless skin, airbrushed, symmetrical face, makeup, studio lighting, professional portrait, instagram model, attractive, handsome, beautiful, plastic surgery, botox, 8k, masterpiece.

PS - Don't use "asymmetry" in v18; it will bust all your image generations!!! (asymmetrical, too, most likely)

https://civitai.red/images/133070128

https://civitai.red/images/133070591

https://civitai.red/posts/29058926

Some more random images if anyone cares.

same here

@Cyberdelia No, it’s fixed now, I managed to download it

@Cyberdelia thx <3

@valeforal I kept getting 404 pages yesterday, myself. Haven't gotten anything new today to check it. Just mentioning as I think the site was being.. iffy yesterday.

@Cyberdelia no it started working again thank you!

Because people who use Pony only worry about the pink bits.

What would you recommend?

@Afgmc 你是一个人才

Hi Lavinia, put it early in the prompt, even before you describe something else. Like: standing, leaning back, looking away, ( then camera angle)... There is a link to a prompt guide in the descriptions here. That solved some of my pony issues lol. Have fun. :)