Convrot conversion of qwen3vl 4b. near identical to bf16 in terms of adherence and quality and usually near identical output to bf16. the learned version might slightly outperform the simple. but simple handles the bf16 layers differently and you might prefer that one.
My
Optimized fork of Ai-Toolkit
If you want to support my work I accept XMR monero
865BrcfWLdwELwuq5faV1uVTbh93zVK6AUYLY2c3mX6sFfAGRfS6axe1kBTYYKuM7ccN7zBZDAZvnT7E4NKmUazySdbpc7pDescription
FAQ
Comments (2)
What does this do?
is it uncesored? smarter? faster? made better for x y or z image styles?
Thank you for sharing and not paywalling but what does it actually do?
It's in the description. It's a highly optimized quantization of the text encoder which brings it more or less to parity of the bf16 version, but at half the file size.