Replies: 3 comments
-
If I use the TAESDXL vae, the image it's sharper and has a more detail comparison with the KSampler preview. |
Beta Was this translation helpful? Give feedback.
-
Using the KSampler Advanced with the default SDXL vae, also gives good improvements. The image I want is between these 2 images. Or is there another custom VAE for SDXL, which also preserves the shadow detail ( like this one for SD15 ). |
Beta Was this translation helpful? Give feedback.
-
Well, I guess we can always blend the images ... It could still use some shaperning, but the result is not that bad, considering our starting image. |
Beta Was this translation helpful? Give feedback.
-
All SD models that require less steps can cause burn in, if the settings are not just right.
It's the same for LCM, SDXL Turbo, SDXL Lightning, and even the newer Hyper-SD models.
But they all look good in the KSampler preview ( TAESD ).
The example uses Juggernaut XL 7 with the new Hyper-SD 4 step lora.
https://huggingface.co/ByteDance/Hyper-SD
I used 6 steps, and 100% lora strength.
It also uses IPAdapter FaceID + InstantID
So, there are many settings I can change to improve the image quality, and I can get good images.
But, this always changes the image a little bit.
What if I want this exact image ?
The image on the left is the KSampler preview ( at step 4 ).
It was only upscaled in Comfy UI ( Lanczos interpolation ).
I mean, it's like 90% there !
That's got to be a technique to get the full image quality, right ?
Beta Was this translation helpful? Give feedback.
All reactions