Region LoRAs

Once Regions from Masks has turned your SAM3 detections into numbered regions, the question is which LoRA belongs to which region. The chained way - one Region LoRA node per character, all wired in series - works, but with four characters it's a wall of nodes and a wall of "which region was region 3 again?" Region LoRAs from ComfyUI-NynxzNodes replaces the whole chain with a table on a single node: rows grouped under the region they land on, each row using the same searchable LoRA picker (with previews and bookmarks) as the pack's LoRA Loader.

If the previous node was "one binding per node," this is "all bindings in one place." The workflow stays the same shape - regions wire in, bindings get added, the set carries on to Regions Apply - but everything you might want to compare sits on one card.

What you set

• regions - the region set from Regions from Masks.
• loras - the table itself: add a region, then add LoRA rows under it. Each row is a picker plus a strength. Several rows may share a region, and their deltas add - so you can stack a character LoRA and a clothing LoRA on the same subject, or split a single character across two LoRAs.

The drag interactions are where it shines: drag a row's grip onto another region to move it there, or onto a "+ Region" target to drop it into a fresh region. When SAM3's mask order changes and your character lands on region 4 instead of 2, you move the row rather than rebuild the node.

Region numbering is the same convention as everywhere in the group - the number is the mask's position in the batch - but the table's grouping makes that less of a footgun: you're dragging rows onto visually grouped regions, not typing numbers into a void. (If your bindings arrive from a wire and you need the chainable single form, that's what Region LoRA is for.)

Output is the regions set, now fully bound, ready for Region Backdrop (chain it last) and then Regions Apply to patch the model.

How the pieces fit

The whole Regions group is a pipeline: masks → regions → bindings → apply. None of these nodes touches the conditioning, and each LoRA is kept unmerged and run as a per-token live branch at sample time - that's what keeps character A's delta off character B's tokens, which is the thing merged LoRAs can't do. The table node is pure convenience on top of that mechanism; skip it and chain Region LoRA nodes and you get identical results.

Install & caveats

In Nynxz's Custom Nodes: ComfyUI Manager → search "NynxzNodes" (registry build preferred - a hand clone needs the UI built since web/ is gitignored: pnpm install && pnpm build), or git clone https://github.com/Nynxz/ComfyUI-NynxzNodes. No pip dependencies. The group is Krea 2 only and experimental - Apply refuses non-Krea models loudly. And remember LoHa/LoKr/DoRA can't run as unmerged side branches; stick to plain LoRAs here or the skipped layers will show up as missing effects.