Ponyplex is my first finetune of Pony Diffusion, trying to bring what I've learned from BriXL and XenoDiffusion over on Pony for that extra artistic touch.
This was trained for about 60,000 steps over 4500 pictures of realistic, cartoon and anime cinematic shots and about 1000 concept art pictures with both proper captions and WD14 captions. The latter was a decision taken on my 2nd fine-tune attempt because while Pony gives interesting outputs, it has very poor prompt understanding when it comes to natural language...Prompt carefully!
Don't forget to use CLIP Skip 2 (or -2 for you fellow Comfy users)
Looking forward to seeing what you will all be creating!
Description
FAQ
Comments (10)
Looks cool, thank you!
Use the node "CLIP set last layer" with the value -2
Does not seem to import well in Draw Things, not sure why.
wait how do we do clip skip in comfyui?
Use the node "CLIP set last layer" with the value -2
It's absolutely bonkers what this checkpoint does... Very curious about the training data. I love it! My favourite checkpoint out there, seriously.
Love this model, good job!
amazing work
It generates ugly artifacted plastic images like poorly trained Lora, with Euler sampler effect less ugly but the same very low quality.
I have issues only with this one, what weird parameters does it need.. (ComfyUI).
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.