This is AIO file (Diffusion Model+TextEncoder+VAE) for convenience. Enjoy PonyV7!
Original Resources are below:
Notice:
2025/1/3
Scaled fp8 Diffusion Model + fp16 TextEncoder + fp16 VAE is now available, have fun!
2025/1/2
I finally found the origin of the problem. In fp8 quantization of PonyV7, we need "scaled fp8". Otherwise, it'll break the critical part of the model weight.
2025/12/31
Currently, fp16 AIO has no saturation problem. And, I'll try to recreate correct fp8 AIO.
2025/12/31
I found image saturation problem in fp8 AIO, so I'll upload fixed version for that.
Description
Scaled fp8 Diffusion Model + fp16 TextEncoder + fp16 VAE
Scale Algorithm: AbsMax
Exclude Patterns: bias, t_embedder, init_x_linear, final_linear, positional_encoding, register_tokens, cond_seq_linear
Thank you Anthropic. Claude gave me the conversion script what I wanted ;-)
FAQ
Comments (3)
I think it was a mistake for them to release the new Pony in its current state. I understand the need to be relevant in the public eye for things like this, but, this was a poor showing by any measure. It's a barely functional model that's less responsive than even the earliest versions of Pony6.
Hopefully we'll get a big surprised in 2026!
Maybe not by Zony 8
Is SimpleTuner the only tool available to train Pony V7 loras?, I want to receive replies from anyone who made at least one Pony V7 based lora.
I'm asking this because SimpleTuner is being problematic for me to install and configure.
I already asked this in the Pony V7 original.