Does everything a PonyXL model can, but "nearly" photorealistic!
V4 update, improved faces and lighting trained in.
Workflow I use to create the preview images sdxlpony-face-and-upscale-civitai-metadata
What it does:
If you're familiar with Pony models, they are very specific in their prompting and subject focus. I made this model to work the same way, but get closer to photorealistic. It uses the same prompting and LoRAs that other PonyXL models use.
What it doesn't do:
It's not a general model. Like other PonyXL models, it's more designed to work with booru tags and mostly does images with people in them. If you want to make general images, use a standard SDXL model.
What's coming:
PonyXL is trained on images with "exaggerated" anatomy. So I've purposely kept it from getting too photorealistic for the moment due to uncanny valley issues. As I train it more, I'll bring it closer to photoreal. In the meantime, use prompts such as (realistic photograph, depth of field, bokeh, etc) to help get the photo look.
How to use:
I recommend using the standard SDXL VAE. CFG should range from 7 to 10, higher than other models. You can experiment with different samplers. I prefer DPM++ 2S a Karras for consistency and DPM++ 3M SDE Exponential to get wild. Prompting for quality should use the score system like other PonyXL models. (score_9, score_8_up, score_7_up) in the positive prompt and (score_3_up, score_4_up, score_5_up) in the negative prompt. Otherwise your images will be very plain.
Alternately, you can use the Pony PDXL embeddings that I created for much easier use.
Description
FAQ
Comments (15)
Is there a particular reason not to bake the vae in? For fooocus users it requires a know-how to use a checkpoint without it, and even then it's a nuisance and produces not really perfect results.
I'm a fooocus user and the trick i use is adding a refiner either any sdxl (or even a sd1.5) for just the last 1% of the render and it "kinda" acts like a vae without changing the altering the image much. It's of course not ideal but if anyone had no alternative... yeah.
Vae not being baked is great for a lot of users, esp when merging checkpoints - a shame fooocus doesn't let you load any Vae - as baked Vaes are a big no no for me as it bloats the checkpoint for not reason (for a ComfyUI user).
Maybe Fooocus should fix this (VAE selection has been common in nearly all tools for a long while now) vs. everyone else working around it? They'll presumably be done sooner than everyone who creates a checkpoint can be informed and take appropriate steps.
Baked vaes are horrible. I know its not a you issue but its a Fooocus dev issue and they need to sort out vae selection in Fooocus cause forcing vaes on people in models is peak dumb.
@spaq yes, and additionally to that there is an option in the most advanced settings section to make refiner act as VAE only. If I do that and use a 1.5 checkpoint as refiner, colors are more or less ok, but gens still seem somewhat blurry.
As a workaround, you can bake one in yourself in auto1111 if you select model for model A, type a model name, put slider to 0, choose "no interpolation", choose the vae and click merge. I like to bake in juggernaut v9 vae sometimes, or encode one in auto1111 using an image and encode vae node, and make one if the subject I'm trying to make. Increases likeness, especially for dreambooth models.
@user1234123 this never crossed my mind, but sounds like a great idea! Thank you!
I was really wondering for the last few weeks if someone would attempt a photo/realistic/alternative model trained on top of pony because its seems to have a successful rate of anatomically correct (although) exaggerated/anime/furry characters. Thank you for adding a freshly trained model on top of PDXL. Thank your for your work. I'll be trying it as soon as I can.
Interesting model (V1). But problem is with faces. There is nothing to do without face restore.
A lot of loras seem to help the face.
Is it just me, or does this model really like giving a watermark in the bottom right corner? Not a specific mark or anything - just up to three lines of white almost-text.
Yeah, my trick is to upscale about 15% too big and then center crop.
This model is very good!
Would love to see the uncanny valley photorealistic ;). I've been trying with my own merges but I just can't get Pony to merge well.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.