Wan2.2 I2V Lora Loader By Text (HIGH)

Wan 2.2 isn't one model, it's two: a high-noise pass that handles motion and scene composition, and a low-noise pass that refines detail. That architecture is where the "HIGH" in this node's name comes from, and it's also the source of a lot of confusion - a LoRA file aimed at one pass does the wrong thing if you bolt it onto the other. Wan2.2 I2V Lora Loader By Text (HIGH) is the illumorae pack's answer: it loads a Wan 2.2 I2V HIGH LoRA by fuzzy-matching its name from the <lora:...> syntax tags in your prompt text, and applies it to the high-noise model.

Why load by text instead of a dropdown? Because the pack is built for randomization and variation exploration. You write <lora:char-style:0.8> into a prompt string once, and the loader finds the actual file, so you can shuffle, multiply, or swap lora names in text without rewiring the graph. No list to scroll, no dropdown to change per run.

How it works

The node takes a text input, parses every <lora:name:strength> tag in it, then searches your loras folders (recursively, through ComfyUI's standard folder_paths.get_folder_paths("loras")) for a file whose name fuzzy-matches. It filters by two things at once:

• Variant: the node only accepts HIGH files. The HIGH markers it recognizes are highnoise, highres, highfreq, high_noise, high, and hn (as whole words or suffix patterns). If a matching file has neither HIGH nor LOW in its name, it won't be picked up here.
• Model type: model_type_mode (I2V_ONLY default, T2V_ONLY, INCLUSIVE, NONE) filters on whether the filename carries an I2V or T2V marker. I2V_ONLY means "don't grab that T2V lora by accident."

Then it applies the lora to the model and clip inputs and returns them - plus a lora_info string so you can see what it actually matched.

The one input you'll actually touch

clip_strength_override (default -1). The semantics here matter: Wan 2.2 HIGH loras typically affect the model only, so the default clip strength is 0 - the tag in your prompt still says <lora:name:0.8>, but that strength applies to the model, not the text encoder. If you want the lora to also bend the CLIP, set clip_strength_override to a real value (0–2) and it replaces the default. This is the subtle bit that trips people who expect the tag's strength to be universal.

Outputs: model, clip, lora_info - the first two wire straight into the high-noise sampler leg.

Installing it

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

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

Restart. No models to download; the loras themselves are your own files in models/loras.

Troubleshooting

The classic failure is "no match found" - check the console with DEBUG_MODE on (optional boolean input) and it'll print the search pattern, candidate count, and best score so you can see why it missed. Usually the answer is your lora file name doesn't carry a recognizable HIGH marker, or it has a T2V marker while you're in I2V_ONLY. And remember the speed-lora lesson the community hammered out: if your clip looks plasticky, check you didn't accidentally put a speed lora on this pass. The high-noise pass is where composition and motion live; keep it clean, and use the LOW loader for the detail pass - that division of labor is exactly why these nodes exist in a pair.