do you need a special VAE loader to use in comfyui?
i tried it but did not work.
ditto
It needed conversion, here's my quick one that does work in Comfy: https://huggingface.co/Kijai/QwenImage_experimental/blob/main/e2e-qwenimage-vae_comfy_fp32.safetensors
Thanks, I only got distorted images with it anyway. Your worked burtresults bad. Not sure what I do wrong :)
I suppose it won't work with existing Qwen models? It does work for encode-decode test:
hi thanks for your interest. As
@Kijai
mentioned, the new e2e-vae's wont work directly with existing qwen models without additional finetuning. Please find the details for the new release here:
https://end2end-diffusion.github.io/repa-e-t2i and https://github.com/End2End-Diffusion/REPA-E
Does that mean that a lora trained using base qwen image model and e2e-vae won't benefit from reduced training time?
And consequently, the obtained lora would not work well with base qwen image model?
Ideally that will require finetuning the original model to adapt to the e2e-vae first. For now we mostly explore use of e2e-vae's for training from scratch. we are actively working on releasing an adapted version of qwen and flux models w/ e2e vae's.