LoRA Optimizer (Legacy)

Stack two or three LoRAs together - a style, a character, maybe a lighting LoRA - and you'll eventually hit the thing everyone hits: oversaturated colors, weird artifacts, or a face that looks subtly wrong in a way no single LoRA does on its own. It's a known, common complaint, not a you-problem. The usual fix is trial-and-error strength tweaking. LoRA Optimizer does something more useful: it actually looks at where your LoRAs disagree, layer by layer, and picks a sensible way to combine them instead of you guessing.

The "(Legacy)" in the display name is worth explaining up front. The pack now ships a newer, modular Optimizer + Settings-node combo, and this node is the older all-in-one version where every parameter - sparsification, merge quality, architecture preset, the works - lives directly on the node instead of behind a separate Settings node. It's not deprecated or broken; it's simply the version that predates the split, and it's still the one the pack's AutoTuner-bridge workflow requires. If you just want to point LoRAs at it and get a good merge without thinking about the settings split, this is the one to use.

How it works

The node runs a two-pass streaming pipeline. Pass 1 aggregates conflict and magnitude statistics per weight group and discards the raw diffs, keeping only lightweight scalars. Pass 2 recomputes each group's diff and decides a strategy for it specifically: a group only one LoRA touches gets weighted_sum at full strength; low-conflict overlapping groups get weighted_average; genuinely conflicting groups get TIES-Merging (trim, elect sign, disjoint merge), which resolves the disagreement instead of averaging it into mush. That per-group granularity is the actual insight - a face LoRA and a style LoRA might only clash in a handful of attention layers, and blanket-averaging the whole model would waste correction on the 90% that never conflicted.

The inputs and outputs that matter

• model and lora_stack (required) - your base checkpoint and a LoRA Stack node's output. This is the list of LoRAs actually being merged.
• output_strength (default 1.0) - master volume on the merged result. Set it to -1 for auto, which lets the optimizer pick a strength that compensates for energy naturally lost during merging - the easiest way to stop fiddling with this by hand.
• clip (optional) - plug in your text encoder so merged LoRAs also affect prompt understanding; leave it disconnected for video models that don't use CLIP.
• auto_strength (default enabled) - the setting doing the most work behind the scenes. It automatically turns down individual LoRA strengths before merging so 3+ stacked LoRAs at full strength don't blow out into an oversaturated mess. Leave it on unless you have a specific reason not to.
• optimization_mode (default per_prefix) - how strategies get picked. additive (simple weighted addition, no conflict resolution) is the one to switch to for edit, distillation, or DPO LoRAs - more on why below.
• Everything else - sparsification, merge_refinement, strategy_set, decision_smoothing - is quality/interference tuning for harder merges; the defaults are sensible, and the companion LoRA Merge Settings node covers the rescue knobs if a merge still comes out wrong.

Outputs: model/clip wire straight into your sampler. analysis_report (a STRING) is genuinely worth connecting to a Show Text node the first time you use this - it prints exactly which strategy each layer got and why. lora_data feeds Save Merged LoRA if you want the result as a portable file. tuner_data and settings_source exist for chaining with the LoRA AutoTuner - run the AutoTuner to find the best config, then switch this node to manual and tweak from its recommendation.

Installing it

ComfyUI Manager: search "LoRA Optimizer," install, restart. Manual:

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

Restart ComfyUI - nodes appear under loaders. No models to download and no unusual dependencies; it runs on the tensors your LoRA Stack already loaded.

Common issues & troubleshooting

Don't put distillation, edit, or DPO LoRAs in the stack. This is the pack's own explicit warning: distillation LoRAs like LCM, Lightning, Turbo, and Hyper carry precisely calibrated weights that a conflict-aware merge can quietly break. Apply those through a normal Load LoRA node upstream, and feed only style/character LoRAs into the optimizer. If you must include one, switch optimization_mode to additive and disable sparsification so nothing gets trimmed out from under it.

It only sees LoRAs in its own stack. If your workflow already runs LoRAs through regular Load LoRA nodes upstream, this optimizer has no visibility into them - those patches just stack additively on top of whatever it merges. The pack ships a separate LoRA Optimizer (Inline Chain) node for exactly that situation; it isn't this one.

The merge is still oversaturated after all this. Check auto_strength is actually enabled first - it's the single biggest lever. If it's on and the result is still fried, that's what the tame_layers setting on the LoRA Merge Settings node exists for.

Running out of VRAM on merge_refinement=full. The SVD alignment step (KnOTS) uses more VRAM than the default; drop to refine or none, or set svd_device to cpu.