What's the use case for the "Discard weights with matching name" in the model merge tab?

[WildBanjos]

So, apparently in commit 112416d an option was added in the checkpoint merger tab to remove weights with matching name. Just curious what the intended use case is or if anyone else has experimented with this yet.

[mcmonkey4eva]

Very late response, but for anyone finding this later - it's to discard NonEMA weights AKA prune the model. (Or as the label mistakenly suggests, you can discard the EMA weights and prune the wrong half of the model I guess).

(NonEMA = Training weights, EMA = Inference weights. In practice only the inference weights matter as training-resuming never happens. Also in practice the difference barely matters as EMA doesn't change much in diffusion models).

[zethfoxster]

i feel like i understand it less now that its been explained

[mcmonkey4eva]

@zethfoxster This is a historical post from early 2023 relating to a tool in auto webui meant to correct issues in the original SDv1 release models, and no has no relevance to the modern world