VividPDXL Real (Near photorealism)
Why use this? It has decent color, skin texture and eye quality for a pony model, and can achieve photorealism.
NO UNNECESSARILY LONG NEGATIVE PROMPTS, MOST OF THE TIME THEY HURT RATHER THAN HELP!!!
My usual visual tags are volumetric lighting, depth of field, cinematic, vibrant, shadows, detailed but these may force realistic images even if you are trying to do a different style.
Use source_photo and Realistic tags to help get realistic images.
Make sure to keep CFG around 7 and Sampling Steps around 30.
Always use score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up you can remove the 4,5 and 6 one and you likely wont have any issues.
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, just describe what you want, tag1, tag2Recommended Negatives: artist name, signature, only add more if you want to avoid things appearing in your generation like styles or objects/people. You don't need to use tons of random negatives like deformed fingers I highly doubt they work.
Use source_anime, source_photo, source_3d, etc.. to force or avoid certain styles. You won't see the style change that much but it can help.
If you really want styles other than photorealism, you can try, but its not guaranteed. Try using words like style of "stylehere", while putting source_photo, realistic in the negative prompt as well as describing elements of that style. Certain visual positive prompts can also force a realistic style as well, so try experimenting.
Check out this extension(tagcomplete) for automatic111 users to help with finding tags for pony models.
Also use danbooru to help find tags, it has a tag count to show you how popular certain tags are.
VividPDXL V1.1 (2.5d realistic style)
Why use this? This version produces a more 2.5d like style compared to REAL.
A slight update over 1.0 improving color slightly, and detail.
My usual settings for pony models are as follows:
Res: 1024x1024
Sampler: DPM++ 2M Karras (DPM++SDE Karras, DPM++ 3M SDE Karras are all excellent as well)
Euler for lower detail images.
Steps: 30
CFG: 7
Hires Denoising: .3 - .4
REAL
This update can create photorealistic images, but can sometimes creates a 3d look depending on prompts and loras.
V1.1
Slightly improved color/contrast, maybe a slight improvement to detail and merged the sdxl_vae.
V1.0
Troubleshooting:
If you see a faded image with white spots you need to download sdxl_vae and place it in your models/vae folder.
REAL is a merge of a high contrast 1.1 and Zonkey
V1.1 is a merge of...
My Model Vividvisions
Which is a merge of Duchaiten-NijiUncen v1Lightning and JuggernautXL V9RDPhotoLightning
Description
FAQ
Comments (37)
Make this available on civtai generator?
Civitai generation services are selected already in permissions, but it never became available for whatever reason. I have to use the same permission settings as base Ponyxl .
@insertusername How strange! :(
@lowkey_ I checked in with support, looks like its a scale issue, eventually this model will hopefully become available. This is what they said. "We've outstripped our capacity to host all the models in our GPU datacenter, and we're steading sourcing new cards to be able to handle the load."
@insertusername ooh! too bad, cant wait to test it out :)
indeed our inference providers are not capable of providing enough space to allow for all checkpoints to be used online. for the moment we cycle through all checkpoints from time to time, eventually our goal is of course to offer online generation access to all checkpoints via our own datacenter. due to the still very tough situation on the gpu market this takes way longer than initially expected (perhaps hoped), though.
@eurotaku please cycle this one in, there are many ponyXL checkpoints that could use a pause :D
Pretty nifty checkpoint, thank you.
Of all of then (realistic ones) this one has the best prompt understanding and skin texture quality, I can (almost) make a hentai pose with minimal effort.
Nice! Almost realistic and actually listens to the prompt. Seams to work with the same prompts I've used for other Pony models but this one looks much better
Glad you like it. If you want want a more realistic look make sure to try the VividPDXL Real version.
@insertusername forgot to reply, yes the VividPDXL Real version is very awesome! I think it's my new favourite checkpoint, been using it a bunch the past few days. Thanks for letting me know since I did forget to check for the newest version
Wow crazy good. Pretty much does anything you ask of it!
Crazy good!
I tried to train a realistic female face with this checkpoint with 3480 steps, but the result is far from what I expected, the person doesnt look alike. Were my steps not high enough?
From looking at the celebrity loras for pony, some people used a very large amount of steps, there's one guy that has posted a massive amount of loras but doesn't post his steps unfortunately. I don't have much experience with that either. However I just started two test loras on Emma Stone with this model, so I'll let you know in a bit if I have any suggestions.
Edit: I got ok results with a 2000 step run with a .8 strength and less than 20 images. I'll DM you.
@insertusername +1 also curious!
Can you please make it available for on site generation?
I wish I could make it available. I've tried, It's on civitai's end at this point I believe.
@insertusername that sucks, there's only one realistic pony model available , pny, but it doesn't do vivid and detailed photos all that good. I wonder if you contacted them directly you'd have better luck. I think you can open a support ticket.
It's not like they aren't against NSFW either,they have most nsfw loras listed along with the vanilla pony and fae pony.
My PC just can't do sdxl sadly. I'm stuck with 1.5, or using the site for now.
@Taco360 if you can do 1.5 you should be able to do sdxl. Forge and comfyui have performance improvements that make it viable. There may be others, but I don't really pay attention to them enough to know what their performance is like. And yeah I did contact them for info, You can see another comment on this page to see what I mean. All pony models are heavily nsfw btw, there is not a huge difference between them prompt wise.
@insertusername I've tried it with comyfyui, sadly my computer always freezes when doing sdxl.
I think they struggle with loading time of different models, not sure what infra they have but looks like hdd on prem, cause it takes ages to load model there, if it's cloud then +network time/cost, hope at some point they will be able to sort it out
this release is sick!!! Thanks for the prompting help too! Love the fact that we can keep the prompts short and sweet!!! Amazing
Glad you like it, prompting for all ponyxl models is very similar however. You should start every prompt using any model as short as possible then slowly increase the size as you add more details.
@insertusername yeah well, i love it! Keep up the good work! This is one of the new stars on civit!! And it came out really good, perfectly cooked!
This must be the most enjoyable model I've tried for a long while, I'm having a lot of fun with it!
The model does not support viewing angles well: top view, bottom view. Will there be any fixes?
I've never really had trouble with those viewing angles, you need to use danbooru style tags like "from above", "from below", "from side" etc.
Been using this model exclusively lately for most of my gens. Probably the best for photorealistic fantasy currently! Also does muscled bodies really well. Looking forward to more.
Please make this available for on site generation.
have you considered training REAL on the XL refiner model too? That would be a full kit of awesome OoO
My favorite Pony model! Looking forward to future updates on it!
⠀⠀Let's compare some PONY models 🦄
⠀⠀⠀⠀🌍 See here my detailed test 🌎
Currently my favourite and go to model for anything realism fantasy! Can be combined with different sources like photos, anime, and furry. Best part? 9 out of 10 times eyes, hands, and overall proportions turn out almost perfect at the first tries!
Will this have on site generation eventually? it looks pretty nice.
Its always been on civitai to have it on site generation, at this point I doubt it. They mentioned their servers were overloaded last time I asked.