LoRA Tag Power Loader

This node exists because of Wan 2.2. Alibaba's video model denoises in two passes - a high-noise expert that handles motion and composition, a low-noise expert that refines detail - and the community quickly figured out you often want different LoRA strengths in each. The standing trick from the Wan scene: apply your speed LoRA to the low-noise pass only, and leave the high-noise pass clean. The LoRA Tag Power Loader is the node that makes that trick a tag in your prompt instead of a tangle of stacked loaders.

Mechanically, it's a text-driven LoRA loader in the spirit of rgthree's Power Lora Loader and ImpactWildcardEncode. You embed <lora:...> tags directly in your prompt, the node parses them out, applies the LoRAs with ComfyUI's universal load_lora_for_models() (the same code path stock loaders use, so it works across SDXL, Flux, Wan, Qwen, and Z-Image), and hands you the cleaned text plus the patched model and CLIP. Unlimited LoRAs from one text box, in-memory caching of loaded files, and only the first occurrence of a duplicated tag gets applied - the rest are logged and skipped.

The tag syntax and the one gotcha

Three formats, from simple to full control:

• <lora:name:0.8> - one weight for both noise passes.
• <lora:name:1.2:0.3> - different weights for high vs low noise.
• <lora:name:1.0:0.8:0.6> - high, low, and an explicit CLIP weight.

Here's the gotcha, and the README's examples won't tell you: a two-part tag sets the model weight for both passes but leaves CLIP on default_weight - the code does that explicitly. So <lora:style:0.8> with the default default_weight of 1.0 gives you M:0.80 C:1.00, not C:0.80. If your conditioning feels off, set default_weight to match your typical weight, or use the four-part form to pin CLIP explicitly.

The inputs and outputs that matter

Only a handful you'll actually touch. text is the whole point - paste your prompt with tags inline. noise_mode picks which weights get used: high_noise, low_noise, or auto (which averages the two). model and clip are both optional - leave clip unplugged for model-only loading. Then three quality-of-life knobs: weight_multiplier (a global scale on every tag, handy for A/B-testing strength without editing text), video_model_mode (enables WanVideo/Hunyuan LoRA key standardization for AI Toolkit/LyCORIS/Diffusers-format files), and auto_trigger (reads trigger words from the LoRA's safetensors metadata and injects them where the tag was, so "portrait lora:anime_style:0.8" becomes "portrait anime, detailed rendering").

Five outputs, and the last three are what make it nice: MODEL and CLIP go to your sampler/conditioning like any loader, text is the cleaned prompt for CLIP encoding, lora_info is a human-readable list of what actually loaded with final weights (read it when something looks wrong), and trigger_words is a comma-separated string of every trigger found, handy for debugging or building a prompt manually.

Installing it

Standard custom-node install - no requirements.txt, no model downloads, nothing heavy:

cd ComfyUI/custom_nodes
git clone https://github.com/jonstreeter/comfyui-Lora-Tag-Power-Loader

Then restart ComfyUI and look under loaders > LoRA Tag Power Loader (also searchable via ComfyUI Manager by the pack title). The only optional dependency is ComfyUI-WanVideoWrapper, which video_model_mode reaches for to standardize video-LoRA keys; if it isn't installed the node logs a fallback to standard loading and keeps going. A small wink at the README: its install snippet literally says git clone https://github.com/yourusername/... - use the real URL above.

Where people get burned

• "LoRA not found" - the matcher tries exact, stem, prefix, then case-insensitive, so <lora:my_style> finds my_style.safetensors, my_style_v2.safetensors, even subfolder paths starting with my_style. If it still misses, check the console for the specific file it looked for and make sure the file's in ComfyUI/models/loras/.
• Video LoRAs doing nothing - you need video_model_mode ON, and ideally WanVideoWrapper installed; without it, non-standard key formats silently fall back to plain loading.
• Weights not matching the tag - see the default_weight gotcha above, and read the lora_info output: it shows the exact applied M/C strengths per tag, which is the fastest way to tell whether the node or your math is wrong.

Honest verdict: this is a young, single-maintainer pack with almost no community footprint yet, so treat it as a tool, not a creed. For a plain SDXL or Flux workflow, Power Lora Loader or the stock loader will do everything you need. Where this earns its place is the Wan 2.2 dual-pass workflow: run two instances with the same text - one high_noise, one low_noise - and feed each into its corresponding denoise pass. That's a genuinely awkward thing to do with any other loader, and this makes it copy-paste simple.