LoRA Merge Settings

This node doesn't merge anything. It's a bundle of advanced knobs that plugs into the merge_settings input on an Optimizer Settings or AutoTuner Settings node - a way to hand a pile of fine-tuning options to the actual merge nodes without cluttering them. If you've never seen this node before, that's normal: the defaults are good, and most people should never need it. It exists for the specific moment your merged LoRA stack comes out looking wrong and you want to know which lever to pull.

Some context on why that moment happens at all: stacking LoRAs is famously unpredictable - interactions between two or three style/character LoRAs can produce oversaturation or artifacts that neither one shows alone. The LoRA Optimizer's whole job is figuring out, per layer, whether your LoRAs agree or fight and picking a merge strategy accordingly. Most of the time its automatic choices are right. This node is where you go when they aren't quite.

What each setting actually does

The bundle covers three rough categories. Leave-alone-unless-broken: normalize_keys (remaps LoRA keys from different trainers like Kohya/AI-Toolkit/PEFT to a common format so mixed-trainer stacks merge correctly - keep it enabled) and architecture_preset (auto-detects SD/SDXL vs. DiT vs. LLM-style thresholds; only override if auto-detection is visibly wrong for your model).

Stability tuning, for when re-running the same stack gives you inconsistent-feeling results layer to layer: decision_smoothing (0.25 default) blends each layer's merge-strategy decision toward its neighbors' average so adjacent layers don't flip between "average" and "TIES" on noisy data - 0.2–0.4 is the sane range. smooth_slerp_gate makes the SLERP blending decision use a smoothed per-layer similarity score instead of a single global one; try it if results feel inconsistent between runs.

Rescue knobs, for when the merge is actually fried - oversaturated, blown-out, "burned" the way an overcooked CFG value looks: tame_layers (0 = off) caps individual LoRA layers whose weight delta is disproportionately large relative to the base weight, which is exactly the kind of "hot layer" that pushes activations off-manifold and causes frying; 1.0 fully clamps them. tame_threshold (default 0.3) sets how hot a layer has to be before taming kicks in - lower means more aggressive. star_eta (default 100 = off) is a separate per-LoRA spectral-cleaning pass: values below 100 SVD each LoRA layer and keep only the singular components carrying that percentage of the signal, which helps when you're merging many or heavily conflicting LoRAs but will just discard fine detail on a clean 2–3-LoRA merge, so leave it at 100 unless you're in "many conflicting LoRAs" territory. Rounding out the memory side: vram_budget and cache_patches control where merge results live (system RAM vs. GPU) - the same tradeoff you'll see on the main Optimizer node.

auto_strength_floor and auto_strength_floor_mode deserve a specific mention: auto-strength (on the Optimizer itself) automatically turns down individual LoRA strengths to stop combined stacks from overexposing, and the floor here is how far it's allowed to shrink them - auto picks an architecture-aware default (motion-heavy video models get a gentler floor than image models), manual lets you set the floor slider yourself, down to 0 (no limit) or up to 1.0 (never shrink at all).

Output

A single MERGE_SETTINGS object - wire it into an Optimizer or AutoTuner Settings node's merge_settings input. It doesn't connect to a model, sampler, or LoRA stack directly.

Installing it

Same pack as everything else here. Easiest path: open ComfyUI Manager, search "LoRA Optimizer," install, restart. Manual install if you prefer:

cd ComfyUI/custom_nodes/
git clone https://github.com/ethanfel/ComfyUI-LoRA-Optimizer.git

Restart ComfyUI afterward - nodes show up under the loaders category, and this node specifically lives in the same node picker group as the rest of the LoRA Optimizer family. No model downloads, no extra dependencies to worry about here.

When to actually reach for this node

Skip it entirely on a first pass - build your stack, connect the plain Optimizer, look at the result. Come back to this node only if: the merge looks oversaturated or "fried" (try tame_layers first), you're merging four or more LoRAs and it feels like they're stepping on each other (star_eta below 100), or results feel unstable between otherwise-identical runs (decision_smoothing, smooth_slerp_gate). If none of that describes your problem, the widgets on the Optimizer node itself - auto_strength, sparsification, merge_refinement - are more likely to be the fix, and this node isn't the first place to look.