Problems with artifacts and low quality
UI: Forge
Checkpoint: AutismMix SDXL
CFG: 1
Steps: 8
LoRA: sdxl_lightning_8step_lora.safetensors
Sampler: Euler A SGMUniform
When I generate an image with CFG 1 the image quality is very poor, opaque, which doesn't happen when I increase it to 3, it's very close to a generation with normal settings, but it ends up causing some artifacts, in theory even with CFG 1 shouldn't the generation be close to the quality of a normal generation (Euler A/25 steps/CFG 7)? And by upscaling with img2img some artifacts disappear, but those that remain are "enhanced".
Any solutions?
CFG: 1
CFG:3
Don't use Euler A, use Euler. CFG=1.
Try out the official workflow json so there is no mistake.
Don't use Euler A, use Euler. CFG=1.
Try out the official workflow json so there is no mistake.
I've done what you asked, but the images are still opaque.
And I've lost the traditional anime look, what merge has.
I am guessing if it is not setting issue, then maybe this particular stylization model are less compatible with SDXL-Lightning.
I am guessing if it is not setting issue, then maybe this particular stylization model are less compatible with SDXL-Lightning.
Doing other tests, I got great quality using DPM++ 2M SDE SGMUniform with CFG 2 (no major artifact problems), it's pretty close to what the model can originally do. But it's a shame that this method doesn't seem to work as it should with this model. But this method (SDXL Lightning) is very good, maybe it will improve in the future. Thank you.