Fantastic Lora Mimic 🪞

Here's the situation the Fantastic Lora Mimic solves: you've got a LoRA stack lovingly arranged in an rgthree Power Lora Loader, and you want that same stack applied to a second model path - say the low-noise pass of a Wan 2.2 split workflow. The obvious move is to tap the source loader's already-patched MODEL and feed it onward. But sometimes you can't, or shouldn't: the source's model is already modified, the graph routes differently, or you simply don't want a wire threading through the whole canvas. The Mimic applies the source's LoRA list to its own model/clip - clean, independent, no inherited patching.

Straight up, the README calls this experimental, and you should believe it. The Mimic reads other nodes' configured LoRAs by inspecting the graph through informal ComfyUI frontend internals rather than any official interface. It works, it covers the loaders people actually use, and it can break with a ComfyUI update or with a loader it has no adapter for. Treat it as a convenience, verify it mirrored what you expected, and don't build a production pipeline on it yet.

Two ways to feed it

1. Pick (recommended). With nothing wired, choose a source node from a dropdown - it mirrors that node's LoRAs live, read from the graph in the frontend before execution. The source dropdown lists compatible nodes by #id title, with (auto-detect) when there's exactly one, or (none). Supported families: the pack's own stack nodes (they carry a lora_data blob), the stock LoraLoader / LoraLoaderModelOnly (and shape-compatible forks like pysssss's), rgthree's Power Lora Loader, and numbered-widget stackers like Efficiency's LoRA Stacker and Comfyroll's CR LoRA Stack. This path gives you per-LoRA control: each mirrored LoRA can link (its strength tracks the source live - set it once on the source, forget it) or be unlinked so you can override its strength on the Mimic independently. live_mirror is on by default; turn it off to only update on demand, or hit ↻ Pull now.

2. Wire. Connect any LORA_STACK output into the lora_stack input - the pack's own loader emits one, and so do Efficiency-style stackers. The tradeoff: a wired stack is just resolved tuples computed at runtime, with no per-LoRA link metadata, so you lose the strength controls - it applies the stack flat. To change a strength, change it upstream.

The High/Low mode (Wan 2.2 and friends)

Split models like Wan 2.2 use a high-noise and a low-noise pass, and their LoRAs usually ship as pairs (coollora_ep19_high / coollora_ep234_low). Flip the High / Low Model Mode switch and each mirrored LoRA gains companion controls: 🔎 find companion searches for the closest-matching name for the other half (ignoring the noise token and volatile epoch/step numbers, so _ep19_high still finds _ep234_low), with a low/high tag to keep it straight. The companion replaces the original on this model and gets its own strength; you can also also apply original to stack the source LoRA on top (handy for a shared speed-up LoRA identical on both halves), or use source lora when there's genuinely no counterpart. Nothing silently drops - no companion chosen and "use source lora" off falls back to the original.

Inputs, outputs

Required: model and the hidden lora_data blob; optional clip and lora_stack. Outputs: MODEL and CLIP (the patched results to send downstream), lora_stack (the resolved stack, re-emitted so you can chain), and mimicked (a STRING summary of exactly what got applied - read it after your first run to confirm the mirror worked).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Adudeguyman/comfyui_fantastic-loras

or ComfyUI Manager → search Fantastic Loras, restart, hard-refresh. No dependencies. Both example workflows in the pack are Ideogram4-based, which is the other big split-model target this node exists for.

Common issues

The core one is trust: because it reads graph configuration rather than computed results, it reflects what a source is set to, not anything a node computes at runtime in Python (the pack's own randomizer is fine - the frontend bakes the rolled pick in before queueing). So: check the mimicked output before a long run. If a source isn't showing up in the dropdown, it's not in the supported families - wire a LORA_STACK instead. And if something breaks after a ComfyUI update, this is the first node you should suspect; it's the price of the feature.