Does everything a PonyXL model can, but "nearly" photorealistic!

@Zovya this is the first pony model I try :)

https://civitai.com/models/221845/lowra-offset-noise

@AbundantFather This is for SD 1.5, not XL/Pony

Actually for anyone interested I found a neat solution. I still have to experiment with it more, but so far I have found it can vastly improve images with this model, not only for darker images but also in general if you play with the settings. It's the Auto1111 extension "vectorscope CC", which allows you to control brightness, contrast and saturation. Very simple to use and very useful!

that's not a pony lora though; it doesn't create artifacts merging?

You can apply the Lightning lora to any model. It's also quite easy to merge into a checkpoint with kohya-ss.

@focalbluebell Thanks so much man. I should know this... LOL i've been using comfy long enough. LOL. I have a shitty 8gb card, so running sdxl is painful.. i find turbo and lightning models work so much better, but for some reason I didn't realize I could use the lora on a normal SDXL. Thanks so much!!

lightning lora suprisingly works with it (other pony models besides the autism's lightning embedded which technically not using the lora, doesn't correctly work with lightning lora's) try it

I don't get anything like that. Can you post your settings? Sampler, steps, cfg, resolution etc? Maybe post some examples along with them?

@X0l0t0l ty, for replying, https://files.catbox.moe/7yxulw.png metadata included.

85% of pictures have this type of artifact especially around the feet and the mouth sometimes. I XYZd the same prompt with several pony merges and I liked this model but I constantly get Artifacts (but only with this checkpoint merge, reg pony, confetti, and others all work fine.), I tried different VAEs too.

@vrebakijiji535 how many steps, cfg?

@derpmagician I kinda posted the picture so that you had all of the metadata is included... but ok, 30 steps and 7 cfg.

Interesting Question.

I´ve solved it with this positive Promt:

(18 years old african girl, (dark-skinned female:1.5):1)

Interestingly, the AI sometimes produce strange white skinned tanlines.

To prevent this, I ´ve added "(tan, tanlines:1.5)" to the negative Promt.

So far, it works. Created a Batch of 10 Images, and they are all fine.

No A-Detailer or Inpaint was used.

see Gallery

Cheers,

D.

@Zovya Hm weird, whenever I use the tag futanari, or gynomorph, it just shows a normal woman, even if I add penis and such :( But yeah, still very good for the rest :)

sometimes when you enter some object into the prompt, the model likes to merge the penis and this object and at the output we have a polymorph

NightVision XL ?

@rlewisfr346 I'm unsure if Pony and non-Pony models merge well?

@Zovya Unfortunately its not just the face, its really an anime picture with a hint of a 3D shader, not looking like a photo at all. So inpainting the face wouldn't be enough. I will see if using another XL model as refiner might help. Can't wait for V3, thanks man!

You can try to change some IP-Adapter settings to get closer to what you want. For example you can make it start at 0 and end at 0.5 then from 0.5 to 1 this model will try to make it more realistic. Alsto seems like Instantid works better.

@stevoperic I had the same idea with stop earlier but while the preview comes out as realistic, the last steps then turn it cartoony again. Also I didn't have any luck with InstantID. Besides that it needs two CNs it gives me a oversaturated slightly blurry image that's always a portrait even when I prompt full body image.

IPAdapter in general doesn't work well with Pony based models, I've not tried InstantID with it though, I'll have to give that a go!

@Zovya I'm excited for V3 - do you have an eta for when you think it'll be ready?

Seems to be a problem with the site and pony models in general.

@Taco360 I'm not on site im using local auto1111 unfortunately..

Are you having issues with SDXL not working, or PDXL not working? -- Becuse if your loras are trained on SDXL, they won't work AS well with Pony base models. I use Auto but i'm not a local auto user - i'm one of the nomad dorks since my computer tries to screech to a halt when i threaten it XD

I'm thinking a Sino-Germanic woman of distant Norwegian ancestry.... why won't it do that?!? Ah, it will... "Almond-shape eyes, wide nose, high cheekbones, broad forehead, straight blond hair, brown eyes". Like her but don't want to type it in over and over? Save the image and use it as a template OR make about 50 of her, spin her in every direction, play dress-up, play head-shoulders-knees-and-toes, send her through the Civit Lora maker and choose THIS SPECIFIC checkpoint when asked for an engine (don't just take pony) it costs more V-bucks or whatever we have here... but it is worth it. With a good checkpoint such as this (and it is), you'll find checkpoint-specific loras make a HUGE difference. (THEN, you'll have the fun mixed emotions feeling whenever there is a checkpoint update and you wonder if your old Loras will take a hit in quality! YAY!)

EXACT same issue here... my solution? (Because I'm lame and can't figure out ipadapt (it runs well on apple tablets? j/k) Develop your character in a SOLID base model, refine the crap out of it. Then make about 50 images of your character wearing their various outfits, doing the yoga of various poses, etc. THEN make a checkpoint-specific LORA for your character (for this checkpoint, obvs). Then, switch checkpoints, do some laundry while it finishes loading, and create away. You'll find different checkpoints work better for character development than they do for story-boarding and vice-versa. You can have EXCELLENT at one or the other... but, I've found, only mediocre at both. If you are one of those "high-talent" bastards and hand-draw your stuff, skip step one. Make a damn lora. All the best! ***HUGS***

@Inland389 Hi, thanks for your reply! I am a bit confused. Perhaps can you name very specific models in that process that I would use to get the Everclear PNY model to become controllable?

As far as I understand your description you would make a character not in everclear but in a model like SDXL ("solid base model"?), would create images, then a LoRA for SDXL, and then switch to Everclear PNY and use the SDXL lora?

Or do you mean the other way around, make 50 images with Everclear and then train a LoRA for Everyclear and then switch to something like SDXL?

Or do you mean you would make 50 images with Everclear, move to something like SDXL, train a lora with your evercleaer images, and use SDXL with that style for posing?

Many thanks! I hope you understand where I got confused^^

Could be because Pony v6 was trained on about 2.6 Million images and also probably how they captioned the images as well.
v7 is supposed to be trained on 10 Million so I'm looking forward to that haha

@Rule34Diffusion If anything, we have all polished our vocabulary as we dream up synonymous ways of asking for what we want. "Eating an ice cream"? nope... "Licking an ice cream"? ACK! "Tongue out, ice cream near mouth, pink paste on tongue"? BINGO You wanted sprinkles? Go F* yourself. ;)

This is a valuable comment in that you're not happy with it, but you're still appreciated - Well rounded critiques like this are lacking in the community :P - I don't have the same haze issue, but it could just be me -- I have other reasons why this model has been great for me -- I like both Fae's and this one, and they're both the backbone to Virtual Diffusion :)

No Pony realistic merge is currently "realistic" as in photorealistic - as much as someone prompts for it. The only thing that really works (kinda) is using Everyclear (or my personal favorite: VividPDXL) and switch ON the "Refiner" (in A1111/Forge) with a "Switch at" of about > 0.8. As the refiner model use "AlbedoBaseXL". To get even more out of this you could add ADetailer with the same model (AlbedoBaseXL in the "Use separate checkpoint" setting) and an inpainting strength of about >0.2<0.38...

"...(without specifying the age/characteristics of the subject)..." so you don't tell it what you want then complain about it not giving you what you want? Total legit.

10/10 of current sample images perfectly mature. Seems you have an issue with your prompt, not the checkpoint.

See that great big box with "Negative Prompt" on it? THAT is probably one of the most important areas of ANY Pony checkpoint. When you have a SOLID prompt that you believe should be giving you what you want, start tooling with the negatives. Eyes too big? "Big eyes"... all the women looking too young - "child" ... have no talent - download a 1.5 checkpoint. The original dataset for Pony is BASED on anime... scroll through the images on here... for ANY Pony checkpoint... you'll notice that nearly all have young faces... til you find one that DOESN'T... now click it and read the negative prompt... bingo.

I raise you with:
Not everyone is built like us americans, this model ENTIRELY like every other stable diffusion model with absolute respect sir: Depends on what you're prompting. I accidentally ran that same prompt i used in another semi realistic model, and got a YOUTHFUL face yes - because it was an anime character i ported to a realistic prompt -- No it wasn't a LOLI picture, no there wasn't anything illegal with it, but there is a wide variety of people working to do different things with Pony at different times. To each their own sir, if you don't like this model that's fine - but don't exclaim loudly that it's a Loliification because that's incorrect.

@Kaladae Could be a sampler issue as well...
I always do an xyt plot of the major 10 samplers and you'll be surprised how different the output is...

@m_go Yeah, sampler can make a huge difference. A single seed with different samplers can look drastically different. When prompting for couples I like to always try different samplers because some will only do a single actor regardless how heavily prompted it is (based on how it feels about that seed) but others will add the second actor. If I'm about to give up on a specific seed, run a different sampler and find out it's a golden seed just with Euler A instead of whatever else.

If you are running into issues that you're unable to rectify via prompting just use a different model? I haven't encountered this issue with v3 or v2 using the recommended settings without the "score" portion of the prompt. Give it a shot without the "score" so the model is less inclined towards cartoon/anime body proportions and characters.