pink or green/orange dots in generated images
First, thanks so much for this amazing model! I have a question though:
Is the following effect that I get on nearly all generated images wanted (watermark/fingerprint) or a (possibly avoidable) artifact of teh generation?
Which model / checkpoint are you using?
Is it from diffusers or ComfyUI?
Are you sure you have use the correct sampler setting as in the documentation?
Thanks for the quick response!
I am using the provided sample code for diffusers (stable-diffusion-xl-base-1.0 with the unet.load_state_dict method as provided).
I got these results with sdxl_lightning_4step_unetand 4 inference steps, as well as sdxl_lightning_8step_unetand 8 inference steps accordingly.
What do you mean by sampler? I use EulerDiscreteScheduler with timestep_spacing="trailing" as in the demo code.
guidance_scale is also set to 0 as instructed.
width and height are default (1024 iir. correctly)
It might be related to watermarking problem from diffusers lib, not from the SDXL-Lightning models.
Please check https://github.com/huggingface/diffusers/issues/6292 and https://github.com/huggingface/diffusers/issues/4074
@seravee008 Thanks a lot! That seems to be the case indeed!
Adding
class NoWatermark:
def apply_watermark(self, img):
return img
pipe.watermark = NoWatermark()
solved that problem for me (and got rid of any steganographic tracking - I assume this is still OK with the Stable Diffusion license...)
@PeterL1n - maybe these lines could be added to the example code?
The images I generated do not have this artifact. I am using diffusers==0.26.3. Maybe it is already fixed in the latest version?
Hmmm... Somehow I am stuck with diffusers==0.21.1 as it seems... Then sorry for the false alarm!