IS_CHANGED still needs to be removed, as per my last comment in the first PR. Node application is not on a per-step basis - always returning True from IS_CHANGED will break the proper node execution behavior. The only things in a per step basis would be the code that is being registered on the model patcher.

I think there may be an issue you are experiencing that is being incorrectly dealt with IS_CHANGED. Can you describe what was happening when you didn't have the IS_CHANGED? You may need to instead add a wrapper function to initialize/reset the dictionary you're trying to use to count steps at sample time.

@Kosinkadink The reason we always return True in IS_CHANGED is that we need to track the dict current_step, which should be re-initialized before every run to reset current sampling step to 0. I find it hard to retrieve the current step index from the KSampler node so I track it by implementing this

Gotcha, in that case, what you are looking for is making an OUTER_SAMPLE wrapper, an example of which is done for the easycache_sample_wrapper in EasyCache nodes for core.

IS_CHANGED does not work how you would expect - returning True does NOT make it run every time, and is also not the way to make sure the code runs every time the sampler is ran; it is separate from the sampling process entirely. The return value of IS_CHANGED is like a 'fingerprint' - if this fingerprint is different than the last time the node instance was run, then the execution function of the node is permitted to run. If the fingerprint is the same, the node will NOT run and the cached value will be used instead. Always returning True from IS_CHANGED means the node will run ONLY ONCE the whole time it is on the graph (the fingerprint will always be 'True', and will match 'True' of the previous run). In V3 schema, this function got renamed to fingerprint_inputs to be more clear in its effects. The node running is completely separate from the sampling process.

What you'll want to do is make sure the dictionary you want to access gets reinitialized in the OUTER_SAMPLE wrapper. The function that OUTER_SAMPLE wraps runs every time sampling is initiated in ComfyUI, so it's the perfect place to reset the step value back to zero. This function also encapsulates sampling before CFG, so it won't get screwed up by the amount of conditioning being passed in.