Base Workflow published by ComfyUI is correct, but suggested parameters are horrible, especially for the 2B model. I was able to check some parameters and results are very consistent.
14B model works fine with start step 0 (denoise 1.0) and higher
2B model with denoise 1.0 (start step 0) produces quite unpleasant results. But, setting step to 1 makes a true magic.
My GPU is weak, thus I was looking for a robust minimum - how many steps de we need to get enough details. IMHO 28 real steps is sufficient. In my case it means that I start at step 1 and step count is 29.
Testing showed that you can safely set total resolution at 1.2M. Artistic part of the image is fine up to 1.5M, but setting it above 1.2M ruins texts and some classic artifacts (like 2 tails, for example) appear too.
Description
Cosmos-Predict2



