So, I am currently doing an in-depth analysis how Noob V-Pred models work through the denoise schedule and while it still has some way to go, one thing I have found is that in many cases basically all standard shedulers do not match the schedule the model needs to generate a varied and good image.

The reason seems to be that while starting at very hig initial sigma values (around 4500) the model will only really start to form the general composition and pose at sigma values of 30-20. The initial thousands of sigmas are basically setup for the distribution of colour and brightness. If the generation is started at lower values the lighting and contrast tend to suffer from what I have seen in testing. But before going below a value of 30, all the model really does is shifting large splotches of noise around.

Now, this leads into what could be your issue. The first 25% of the sampling process are importent for the larger composition, like pose and overall arrangement. But with most standard schedulers those initial steps will largely be spend at sigma values above 30. Even the beta sampler in it's preset being one of the closest matches wastes around the first 10% of the sampling processes steps shifting variations of mostly formless noise. Steps in which the model usually would focus on the larger scale of the image. Making large scale composition suffer. Leaving the generation of larger composition in part in later steps where the model would focus on initial shapes and detailes. Leading to a lot of it's capability to create the composition being lost.

But just raising the number of steps to have more of them in the early percents of the smapling process to compensate, will also make the model spend a lot of steps in the low sigma values and is liable to overcook the image.

Karras is one of the worst offenders. Only dipping below a sigma value of 30 almost 2/3 of the way into the sampling process. Only producing very low quality results, if any recognisable image at all, because it basically does nothing for 2/3's of the denoise steps.

One solution I have worked out by using some analysis tools to get a return on expected sigma values from the model in combination with trial and error as well as running generations decoding at each step to get a detailed look at how the generation develops is actually quite simple if it is a problem with the schedule.

And that solution is, if you can adjust the Alpha and Beta values of the Beta-sampling-scheduler, to set them to 1. (In Comfyui the preset there is 0.6 for both) That will make the shedule match the best schedule I have found so far for the Noob V-Pred models I have tested. Not guaranteeing it will work, as I did not test on this model. But for those I have tested it worked quite well and improoved composition, cohesion and even prompt adherence a lot.

@Oybou I read through all of that, and I gotta say--I think you've hit the mark pretty well, here. I've used the beta scheduler with the beta and alpha set to 1, and I have to agree it seemingly improved my images a tad. Thanks for the advice, and if you publish anything about it around make sure to keep me posted! Thanks a bunch.

Also, I will say I've noticed that setting the CFG scale as low as it will go, comfortably, tends to produce more varied and creative images, as well.