🌊 WanVideo Wakawave LoRA Loader

If you've run Wan more than twice, you know the pain this node exists to kill: LoRA stacking. The standard Wan graph wants you juggling one loader per LoRA, hand-typing strength values, and mentally multiplying the influence of four separate sliders. The Wakawave LoRA Loader is a single node where you add as many Wan LoRAs as you want - each with its own strength slider and an on/off checkbox - then save the whole combination as a named preset. It's not a model loader. It's a LoRA manager that hands a tidy bundle downstream.

What it actually does

Under the hood it maintains a lora_bundle - a JSON list of {name, strength, enabled} entries that lives hidden inside the node and is managed entirely through the UI. The node's one output, lora (type WANVIDLORA), is that bundle. Importantly, it does not apply the LoRAs itself. You wire that output into a WanVideo Set LoRAs node (from Kijai's ComfyUI-WanVideoWrapper), which is what actually fuses them onto the checkpoint. Forget that hop and nothing happens - the loader isn't touching your model.

Its party trick is the file-size display. On ComfyUI startup it scans every LoRA directory you have configured - including the ones in extra_model_paths.yaml, not just models/loras - and caches sizes in memory, so the node can show each LoRA's file size plus a header total like "8 LoRAs • 6.9GB" with zero latency. Toggle it with the "Show File Sizes" node setting, or pull the same data via the GET /wanvideo/lora/sizes API endpoint. That's genuinely useful for Wan, where a single rank-256 speed LoRA can be 2.7GB and your combined stack is a real VRAM budget line.

The inputs that matter

The info_schema is refreshingly tiny: one optional input, prev_lora (type WANVIDLORA), and one output, lora. Everything else lives in the node's own UI:

• + Add LoRA - pick from a dropdown of everything in your LoRA directories.
• Strength slider (0–2.0) - per-LoRA, this is where the real work happens.
• Enable/disable checkbox - the fastest A/B test in ComfyUI. Disable two style LoRAs, queue, compare, re-enable.
• 💾 Save / 📂 Load / 🗑 Delete - named presets for combinations you don't want to lose.

prev_lora is how you chain loaders: one loader holding your character LoRAs, its output feeding prev_lora on a second loader holding style LoRAs, and so on, with the final output going to Set LoRAs. It's the same pattern as rgthree's Power Lora Loader, just Wan-specific.

Installing it

Same story as its sibling node in this pack - it's a pure UI utility, no model downloads, no API keys, no Python dependencies worth worrying about. ComfyUI Manager can search "WanVideo Wakawave", but the manual route is bulletproof:

cd ComfyUI/custom_nodes
git clone https://github.com/Wakapedia/ComfyUI-WanVideoWakawave.git
# restart ComfyUI

Then find it under Right-click → WanVideo → Loaders.

Where people get burned

• Forgetting the Set LoRAs node. The loader emits a bundle, not a fused model. If you connect lora straight to a sampler input that doesn't accept WANVIDLORA, nothing loads.
• Order matters. Earlier LoRAs have more influence in Wan. Put your character or style-defining LoRA first, polish LoRAs after.
• Strength discipline. Start at 0.6–0.8 and creep up. At 2.0 you're asking for the LoRA to run the whole show, and Wan will happily let it.
• Speed LoRAs are a trap. lightx2v/CausVid step-distilled LoRAs are the standard acceleration tool, but the community has a well-documented PSA: at high strength they wreck Wan 2.2's motion, lighting, and skin. Keep them at 0.4–0.6, ideally on the low-noise pass only. Good news: Wan 2.1 LoRAs generally carry forward to 2.2, so your library survives the base-model upgrade.

If the file sizes column is empty, restart ComfyUI so the startup cache scan runs - it's built once per server session, with on-demand fallback if something was added mid-session.