LoRA Optimizer Settings

The plain LoRA Optimizer node runs on good defaults on purpose - you're meant to be able to plug in a stack and go. This node is what you reach for once you've hit a case the defaults don't cover: artifacts refuse to clean up, a merge is eating too much VRAM, or you specifically want to force one merge technique over another. Connect it to the Optimizer's (or Inline Chain's) settings input and it takes priority over tuner_data and the built-in defaults.

What it controls, and when to touch it

Most of these are fine left alone. The ones worth understanding:

• auto_strength (default enabled) - automatically turns down per-LoRA strength as you stack more LoRAs, to stop oversaturation before it starts. Keep this on; turning it off means you're back to manually finding a safe combined strength yourself.
• optimization_mode - per_prefix (default) picks a merge method per weight group based on local conflict data, which is the whole point of this pack. global forces one strategy everywhere, which is closer to how older LoRA-merge tools worked. additive is simple weighted addition with no conflict handling at all - reach for this specifically when a LoRA in your stack is a DPO/edit LoRA you can't avoid including, since conflict resolution would otherwise trim weights that LoRA needs intact.
• merge_refinement - none (default, fastest) is fine for most stacks. If you're seeing artifacts or color shifts, try refine first: it adds direction orthogonalization and TALL-mask "selfish weight" protection for minimal extra compute. full adds SVD-based alignment on top, at the cost of more VRAM during the merge - save it for stacks where refine still isn't enough.
• sparsification - off by default. If you enable it, prefer the _conflict variants (dare_conflict, della_conflict) over plain dare/della: the plain versions trim weights everywhere, including places where LoRAs aren't actually fighting, which throws away signal you didn't need to lose. The conflict-aware versions only trim at genuine sign disagreements.
• sparsification_density - how much of each LoRA survives sparsification. 0.7 keeps 70%; lower is more aggressive and risks losing detail.
• dare_dampening - only matters with dare/dare_conflict enabled. Controls how much surviving weights get boosted to compensate for the trim; 0 is the standard rescale, higher values dampen it for a gentler (less noisy) result at low density.
• patch_compression - smart (default) losslessly shrinks the merged output where it can; aggressive compresses everything (slightly lossy on TIES-merged prefixes); disabled skips compression entirely. Leave this on smart unless RAM after the merge is a real constraint.
• svd_device - gpu by default and meaningfully faster. Only switch to cpu if you're hitting out-of-memory errors during compression.
• free_vram_between_passes - off by default. Flip it on if VRAM is genuinely tight on a large model; the speed cost is small.
• strategy_set - full (default) lets the optimizer choose from every merge method including SLERP. no_slerp and basic narrow the pool if you want more predictable, simpler behavior.

The inputs and output

Two optional connections beyond the widgets above: merge_settings to share common settings from a LoRA Merge Settings node instead of using this node's defaults, and merge_strategy_override to force one specific merge method everywhere - this is meant to be fed from a LoRA Conflict Editor's output rather than typed by hand. Output is a single settings (OPTIMIZER_SETTINGS) - connect it to the settings input on LoRA Optimizer or LoRA Optimizer (Inline Chain).

How to install it

ComfyUI Manager: search "LoRA Optimizer", install, restart. Or manually:

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

Restart ComfyUI - nodes appear under loaders. Nothing to download; this node is pure configuration.

Common issues & troubleshooting

You connected it but nothing changed. Double-check the settings input is wired to the actual Optimizer node doing the merging - it's easy to build this on the side and forget the final connection, especially in a busy graph.

Merge is noticeably slower now. merge_refinement=full and enabling sparsification both add real compute. If you don't need the extra quality on a given stack, drop back to none/disabled rather than leaving the heaviest settings on for every run.

Out of VRAM during the merge, not during sampling. That's the compression or refinement pass, not your sampler. Try svd_device=cpu or free_vram_between_passes=enabled before assuming your GPU can't handle the stack at all.