ByteDance

/

Hyper-SD

like 547

Hi~
The difference mainly lies on the number of parameters fine-tuned.
For the UNet fully fine-tuning, our model would get learned better.
While for the LoRA training, only the injected sparse matrix would be updated and fine-tuned.
To clarify, our 1-Step SDXL UNet is NOT merged directly from SDXL-Base and 1-Step LoRA.
Instead, it's fully fine-tuned and re-trained from SDXL-Base. Thanks for your attention❤️!

so its just like sdxl checkpoint
? and not a lora?

@roktimsardar123
We have both 1-Step UNet and LoRA on SDXL.
You can pick as you want.

i mean the unet is a checkpoint?
(sorry im new into this)

@roktimsardar123
Do you mean the "checkpoint" in ComfyUI?
In ComfyUI, only the weights with all UNet, VAE and text encoder are called "checkpoint" and should be placed under models/checkpoints folder.
But during training, saved model weights can all be called checkpoints in practical.
The Hyper-SDXL-1step-Unet.safetensors we provide is a raw UNet checkpoint without VAE and text encoder. It can't be used as a "checkpoint" in ComfyUI. It can be loaded by diffusers integration like pipelines.
While the Hyper-SDXL-1step-Unet-Comfyui.fp16.safetensors we provide is a "checkpoint" in ComfyUI which contains all UNet, VAE and text encoder.