How is context merge supposed to be used properly as intended by the author?

[stephantual]

Happy New Year! 🥇

question about merge.json

If i use two big context nodes, I notice i can use them to overide the settings of one another - very handy.
But when i do the same using 'context merge' in the middle, i noticed it 'favors' the top most context fed into it by order decreasing.
I think that's the purpose.

In the example attached, the conditionning is applied to a latent that never should make it to the 2nd ksampler, yet it does.

Does this mean context merge node somehow 'feed backwards' the information back to the 2nd ksampler? I understand comfy reads 'backwards' but still, this shouldn't work, surely, unless the data flows 'back' to the merge?

[rgthree]

Note, you may have a "bug" in your workflow. The bottom conditioning has the positive connected to both context 3's Positive and Negative inputs, which is why that second image looks a little funky. I'm going to assume you meant to connect that bottom ClipTextEncode to Context 3's negative inputs.

The Context Merge node merges non-null values from multiple contexts favoring latter contexts. In your workflow, you're merging in three contexts that ultimately result in a context with Model, Clip, Vae, Latent, Positive, Negative (meaning, that merged node does have the conditioning). Note that the landscape latent is chosen, because it is coming in from the context in the merge node's second input, overriding the square input in the first context.

Here's an addition to your workflow that shows a new context merge that has the same values as the context merge node (note, I did connect that bottom ClipTextEncode to context 3's negative input, instead of the top ClipTextEncode going to both the positive and negative).

[stephantual]

Understood and thank you so much for replying on NYE. I now 'get' why there's a merge node - the bug was because I rushed to create a little example of what I mean :)

Very useful nodes! Happy New Year!