Thx ^^ U 2 (with regards to the new checkpoint at least lol)

I hope Pony Realism v2.3 gets all the recognition it deserves <3

It was a very pleasant surprise for it to be one of the first things I saw when I came back (along with the new OpenAI model; which is of course disappointingly & rather predictably onsite only.

Yeah, v23 keeps generating grandma with weaponized cheekbones. Hate it.

After a day of investigation, I found that zy_Realism_Enhancer_v1 reacted with the new model, causing the abnormal results. I'll try zy_Realism_Enhancer_v2. Thanks again for the update of this SDXL model.

Got this working by lowering cfg to 2, earlier i was using 4 on V22

No, Base is Pony. Pony is based on XL. But yes, a lot of XL Lora work with Pony. The more stronger dependent on Pony the less well it works. But generally Lora have the same problem if the model is far away from what the Lora trying to achieve.

Lora:
First look for Lora based on this Checkpoint
Second look for Pony Lora
Third look for SDXL Lora.

If you are lucky, some Illustrious Lora work as well. Read the description however before experimenting.
Some Lora are just copies on this site (so you can use either variant).

@ZyloO ah, like magic! It works and no longer 404's! Thanks for the heads up!

Im ngl to you mate, you kinda have to train LoRas for all the different models you use. Pony and many other trained checkpoints vary A LOT from base SDXL, and while they should in theory work with one another, captioning and simply datasets are too far away from each other that you should take them as another model entirely

Pony is based on SDXL. I have seen a rarely used Pony Diffusion checkpoint based on SD1.5 as well (currently not in the generator). I also found a merger from Pony and Illustrious or NoobAI or all three together.

What do you want to do? For mostly realistic images I use one Pony/IL/NAI based model. For others it largely depends on the style you want to achieve.

Pro-Tip: Use the search filter when you search for models.

@Raininosi thanks for the reply, that makes sense since the filesizes are similar. i saw only Pony Diffusion V4 is based on SD1.5. I just want to try out stuff i downloaded (mostly realistic). Since my grafics cards isn' the fastest I prefered Sd1.5 since its more efficient in the workflow, SDXL is a bout 2,5x slower, thought i experimented with Turbo,Hyper,Lightning,LCM and others still not done yet.

@pink0909 Pony doesn't do realism. kids can slap that word in front of a model and the other kids nod their heads, but you want SDXL, Kolors, FLux, even SD3.5/L for that. unless all the people in your highly glossy oversaturated candyland life do in fact look like plastic bratz dolls? then i guess pony will work

@Narz i tried to run SDXL , its very strange it gives me BS quality like a shaken camera image, i tried multiple idea liek no-half, nohalf-vae, different vaes inl the standar,and the fp16 version nothing works to solve the bad quality issue. When i run a Pony model for exampe Cyberrealistic Pony XL it does not give me these quality issues like SDXL models do. I remember having the same issue when i tried out SDXL Lightning,Hyper,Turbo models they all looked like a blown up 512x image not like a usualy goo image.

@pink0909 ah, sorry i can't help you, i use Draw Things for Macs, its differently optimized than most windows clients. but i know XL is definitely trickier, lots to do with the latent space / dimensions. it's worth digging into if you have the patience (even if you'e a fan of the Pony look, which i was being a bit Extra about). i'm sure people using your client can help, i spend all my time on a Drawthings discord server

@pink0909 You should try some models that try to do "realistic" images with the Onsite Generator. I have found very few models that I liked. Pony Realism 2.1 is one of them, Real Dream (also popular) and recently CuteRealism which reacts better to Prompts. I also like Goddess of Realism, Untitled.Pony, Illustrious Realism. I recommend use of Lora, especially for Skin (Real Skin Slider and Skin Color Slider to improve brightness or darkness of skin) as well as Amateur Style Slider. LUT color grading is good as well. All these remove the flash effect a lot of images tend to have.

@Raininosi okthank i wil ltry them out later. I was frustrated and delted my 200gb SDXl models i collected before i knew they where useless to me.Good old torrent habbits lol. Since i have space now i will fill it with ponyXL and illustriuos, Anyway since that SDXL issue anoyed me very much i thought well fk it i will try to throgh some SDXl pngs at SD1.5 (Cyberrealsitc for example) and to my suprise it worked mostly (when i diodnt i was lacking the used loras). I will just use Sd15 + highres fix. I just learned recently how to use it correctly for max quality(dont use latent, instead NMKD or similar). Also in hindsight irewatched some SDXL galleries here on civit and realised most images have this exact same look (but a bit weaker) i was annoyed with. Also most dont understand the difference between pixel resolution and optical resolution. One good way to judge, is to compare the eyelashes details on 2 images you want to compare, secondly the hairdetails and last the skin. Eyelashes seem to be the most difficult thus a good performance indicator. Thats my experience so far.

@Narz ok thank you for your reply. No i am not a Fan of pony , just new to it. I am a Fan of Sd15 ;) since i found out how to use it to get "high" resoloution.

@ZyloO Hang on.... Are you saying I could set up the system to use the PR model as a refiner? Or that I could just use the standard SDXL refiner? (Sorry, still trying to wrap my brain around some concepts that never were clear to me when I started.)