Normalize Model

Fine-tunes of the same base model don't all sit at the same scale. One checkpoint might have beefier weights in the attention blocks, another in the feed-forward layers - and when you merge two models that are speaking different magnitudes, the blend can wobble even though each parent is fine on its own. DM_NormalizeModel is DareMerge's answer: it rescales one model's parameter norms toward another's so the merger is combining like with like.

How it works

Feed it two models, model_a and model_b. It computes the ratio of the norms between corresponding parameters, then scales model_a's weights by that ratio (via model_b's, if you flip magnify to on). The output is a patched clone of model_a - so you chain it in front of a merger node, not instead of one. This isn't a merge, it's a preconditioner.

The method dropdown is where the thinking happens:

• attn_only (default) - only scales Q and K relative to each other, and nothing else. The README's pitch: "You should see no difference in the model's performance, but it might make the merge more stable." That's the one to try first, because it's the least invasive.
• q_norm - scales the whole attention stack (Q, K, and the output projection) using Q as the reference.
• all - normalizes weight+bias pairs across the board, the broadest sweep.
• none - passes model_a through untouched (mostly useful as a baseline when you're testing whether normalization is even doing anything).

Where it fits in a workflow

If you're doing an advanced or DARE merge, the standard chain is: load both checkpoints, run model_a through Normalize Model, then feed the normalized result plus model_b into the merger. The normalization happens once, then you can sweep merge ratios without re-running it. It's especially worth trying when you're merging models that are related but trained differently - the kind of merge that's "mostly fine but the hands are weird" and you're not sure why.

The honest caveats

• The key-group logic is built around SD1.5. The node's layer targeting (generate_key_groups_sd15) explicitly spells out SD1.5's architecture. On SDXL, many of those groups simply won't match and get skipped - so treat this as an SD1.5-first tool and test carefully if you're on SDXL.
• It's subtle by design. If you're expecting a night-and-day change, you'll be disappointed. Normalization is the kind of thing that quietly moves a merge from "almost" to "there," not something that transforms the output. The README says as much - the selling point is stability, not visible difference.
• Magnify is experimental. Flipping it inverts the scaling direction; the README frames the whole normalization approach as "testing out a new method," so expect the rougher edges.

Installing it

One of ~25 nodes in the 54rt1n/ComfyUI-DareMerge pack. Install the pack once via ComfyUI Manager (search "ComfyUI-DareMerge") or:

cd ComfyUI/custom_nodes
git clone https://github.com/54rt1n/ComfyUI-DareMerge

then restart ComfyUI. Dependencies (matplotlib, numpy, torch, pillow) are standard ComfyUI fare; no model files to download.

The bottom line: this is the "try the boring fix first" node. When a merge is unstable and you've already fiddled with ratios, a quick normalization pass is cheap insurance - and attn_only is deliberately safe enough to leave in the pipeline permanently.