Regarding the issue where the generated images in version 5.3 exhibit matrix artifacts
The current version in use is 5.3. When generating images in the 4-step process, there will be matrix-like dot-like artifacts. They are not very obvious but can be easily noticed as there are faint matrix-like artifacts. The symptoms will ease up after the 6-step treatment, but the problem still persists. Sometimes, it is completely absent.
Is there any way to solve this problem?
The sampler uses LCM, while the scheduler uses kl_optimal. The reason why we didn't use the recommended scheduler is that the problems with the recommended scheduler occurred more frequently.
LCM is not a preferred sampler. From my personal experience, I believe the combination of Euler a and Simple is truly perfect. When you're unable to achieve the desired image, increasing the sampling steps and observing the process will reveal the preview image progressively advancing toward your expected outcome. In contrast, LCM typically stops constructing the overall composition around 4-8 steps, focusing solely on refining details. Therefore, I don't consider it a qualified sampler.
@q27916430 I always used lcm/ddim_uniform because I got the most realism, even if it darkened the scene a bit. Do you keep the denoise in the ksampler at 1.0 or do you lower this? Playing around with 4-6 steps and different scheduler/sampler combos, I appreciate the recommendations for others I have not tried.
@q27916430 I always used lcm/ddim_uniform because I got the most realism, even if it darkened the scene a bit. Do you keep the denoise in the ksampler at 1.0 or do you lower this? Playing around with 4-6 steps and different scheduler/sampler combos, I appreciate the recommendations for others I have not tried.
If you find it suitable, then nothing could be better. Nothing in the world matters more than your own feelings. Regarding the issue of high image contrast you encountered, as previously mentioned by another user, reducing the CLIP-Norm to 0.96 might resolve this problem. You could give it a try and see if it works.
@q27916430 I always used lcm/ddim_uniform because I got the most realism, even if it darkened the scene a bit. Do you keep the denoise in the ksampler at 1.0 or do you lower this? Playing around with 4-6 steps and different scheduler/sampler combos, I appreciate the recommendations for others I have not tried.
If you find it suitable, then nothing could be better. Nothing in the world matters more than your own feelings. Regarding the issue of high image contrast you encountered, as previously mentioned by another user, reducing the CLIP-Norm to 0.96 might resolve this problem. You could give it a try and see if it works.
Thanks for the comments. I was using the JSON provided here but don't see anywhere "CLIP-Norm", can you point me in the right direction to modify it?
@q27916430 I always used lcm/ddim_uniform because I got the most realism, even if it darkened the scene a bit. Do you keep the denoise in the ksampler at 1.0 or do you lower this? Playing around with 4-6 steps and different scheduler/sampler combos, I appreciate the recommendations for others I have not tried.
If you find it suitable, then nothing could be better. Nothing in the world matters more than your own feelings. Regarding the issue of high image contrast you encountered, as previously mentioned by another user, reducing the CLIP-Norm to 0.96 might resolve this problem. You could give it a try and see if it works.
Thanks for the comments. I was using the JSON provided here but don't see anywhere "CLIP-Norm", can you point me in the right direction to modify it?
Please add a node named "CLIPNorm" to your workflow. This node belongs to the model circuitry and should be connected in series before the sampler. If the node is not available, you may need to download the corresponding plugin.
The sampler uses LCM, while the scheduler uses kl_optimal. The reason why we didn't use the recommended scheduler is that the problems with the recommended scheduler occurred more frequently.
thanks for the kl_optimal scheduler tip. That solves the dot/bands matrix issue with v5.3 NSFW.
The sampler uses LCM, while the scheduler uses kl_optimal. The reason why we didn't use the recommended scheduler is that the problems with the recommended scheduler occurred more frequently.
thanks for the kl_optimal scheduler tip. That solves the dot/bands matrix issue with v5.3 NSFW.
Sadly it also seems to make the image softer, compared to other schedulers.