Omini Kontext LoRA Merge

Where the LoRA Loader adds one adapter to your pipeline, Omini Kontext LoRA Merge adds several at once and lets you weight them independently. If you've ever wanted your character-insertion LoRA and a painterly style LoRA active in the same generation - or a character LoRA plus an "object placement" LoRA from the README's use-case list - this is the node that composes them.

How it works

It's a thin wrapper over diffusers' pipeline.set_adapters(). You give it a comma-separated list of adapter names and a comma-separated list of matching weights, and it activates all of them on the pipeline with their individual strengths. The combination_type dropdown (linear vs cat) switches how the adapters combine in the PEFT layer - linear is the standard weighted sum you want for stacking, cat is the concatenation mode diffusers supports for some combinations. Default is linear, and that's what you'll use 99% of the time; cat exists because it's a diffusers option, not because it's usually the right call.

The inputs that matter

• pipeline - the pipeline with your adapters already loaded (via one or more LoRA Loader nodes chained in).
• adapter_names - comma-separated, e.g. omini_kontext,watercolor. Must match the adapter_name values you set in the LoRA Loaders.
• adapter_weights - comma-separated floats in the same order, e.g. 1.0,0.6. Length must match the names list or the node raises a ValueError - the one genuinely sharp edge here.
• combination_type - linear or cat.

Output is the merged OMINI_KONTEXT_PIPELINE, ready to feed the generation node.

Where it fits

The natural flow is: loader → LoRA Loader (character) → LoRA Loader (style) → LoRA Merge → Pipeline. Each Loader gives its adapter a distinct name as it loads, then the Merge activates the set. Because it returns a fresh pipeline object, you can also branch - one Merge for the full stack, one Unload for the clean version, and compare both against a fixed seed. The pattern will feel familiar if you've used LoRA stacking in any other Flux workflow; the only quirk here is that it operates on adapter names instead of file names, so getting the names right in the Loaders matters.

Gotchas

The weights list is parsed by splitting on commas - 1.0, 0.6 with a space after the comma actually works because each token gets stripped, but don't push it with fancy syntax. And remember the VRAM reality of this pack: a merged multi-LoRA run of a 12B diffusers model is heavier than a single-LoRA run, and at 16GB you'll feel it sooner. Merging LoRAs is free compute-wise; running the model still isn't. If generation OOMs after you added a second adapter, that's the model bill, not the merge.