Wan2.2 Lora Loader By Text

Wan 2.2 is a two-model model. The high-noise pass composes motion and scene, the low-noise pass refines detail, and most Wan 2.2 LoRAs ship as a pair - MyStyle_HIGH.safetensors and MyStyle_LOW.safetensors - so loading one correctly means loading both onto the right passes. That's a job most lora loaders don't even understand exists. Wan2.2 Lora Loader By Text, from the illumorae pack, is built around it: it parses <lora:...> tags from your prompt text, finds the matching HIGH and LOW files, and applies each to its own model input.

It's the most useful of the pack's three Wan loaders because it's the only one that takes both model_high and model_low inputs and does the pairing for you. The two I2V-only siblings (HIGH and LOW) are single-pass tools; this one is the full two-pass wiring in a single node.

How it works

The matching runs in two phases, per the source docstring. First, a primary match on the base pattern (the lora name with HIGH/LOW keywords stripped), scored at up to 95 for an exact match. Then a pair match: the corresponding file must have the exact same base pattern and be the opposite variant - Style-T2V-HIGH-14B.safetensors pairs with Style-T2V-LOW-14B.safetensors in the same folder. If no pair exists, only the primary is applied; the node logs it, and that's a real case, because some Wan loras genuinely ship as a single file.

fallback_classification (HIGH default, LOW, SKIP) decides what happens to those single-file loras with no HIGH/LOW marker in the name: apply as HIGH, apply as LOW, or skip. The high/low keyword table is the same one the siblings use (highnoise/lowres/hn/ln and friends), and model_type_mode (default NONE here, unlike the I2V nodes' I2V_ONLY) filters by T2V/I2V markers in filenames.

One more piece of behavior worth knowing: the loader tracks which loras it has applied to each model, so you can specify both variants explicitly - <lora:MyStyle_HIGH:0.8> and <lora:MyStyle_LOW:0.5> apply each at its own strength instead of fighting.

Inputs and outputs

• text - your prompt with <lora:name:strength> tags.
• model_high, model_low - the two pass models from your Wan 2.2 loader.
• clip - shared text encoder.
• fallback_classification and model_type_mode as above.

Outputs: model_high, model_low, clip (wire each model to its sampler pass), plus lora_info showing what matched.

Installing it

ComfyUI Manager → search "illumorae" → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/CorvaeOboro/ComfyUI_illumorae

Restart. No downloads - it reads your existing models/loras.

Troubleshooting

If a lora only lands on one pass, check two things: whether the pair file actually exists with the exact same base name, and what fallback_classification is set to. Enable DEBUG_MODE and the console will walk you through the search, the candidates, and which pass got what. And don't sleep on that speed-lora workaround from the community: with this node you can put a speed lora only on the LOW pass (via fallback_classification or an explicit _LOW tag) while the HIGH pass stays clean - the exact setup that keeps Wan 2.2's motion quality while making 4090s practical. That's the killer use case for this node.