WanVideo Lora Select By Name

This is the node behind almost every speed trick and character customization you'll do with Wan in this pack - it's how you load a LoRA file by name from ComfyUI/models/loras. If you're applying LightX2V, CausVid, FusionX, or any character/style LoRA you've trained yourself, it goes through this node.

The one thing worth knowing before you use it

Unlike a typical LoRA loader elsewhere in ComfyUI, lora_name here is a text field, not a dropdown. You type the filename in by hand rather than selecting it from a scanned list. That's a genuine UX quirk worth knowing up front, because it means a typo produces a silent miss rather than an obvious dropdown error - get the filename slightly wrong and you may just get an unmodified base model with no clear signal anything's off.

How chaining works

prev_lora is the mechanism for stacking multiple LoRAs. Rather than a list-based multi-LoRA node, you chain instances of this node together - load your first LoRA, feed its output as prev_lora into a second instance loading your next one, and so on. This matters specifically for Wan's speed-LoRA workflow: the KB's community-tested best practice is applying LightX2V or CausVid to the low-noise pass only, at reduced strength (0.4–0.6), while keeping the high-noise pass clean - full speed-LoRA strength on both passes is what the "speed loras kill everything good about Wan 2.2" warning thread was about. Chaining lets you build that kind of asymmetric setup precisely.

The inputs and output that matter

lora_name (STRING, typed) - the LoRA filename, exactly as it sits in ComfyUI/models/loras. strength (default 1, range −10 to 10) - the usual LoRA weight, with a wider range than most loaders offer, so don't assume 1.0 is a ceiling if you're intentionally overdriving an effect. prev_lora (WANVIDLORA, optional) - chain in a previously loaded LoRA to stack it. blocks (SELECTEDBLOCKS, optional) - wire in a selection from WanVideo Lora Block Edit if you want this LoRA restricted to specific transformer blocks rather than applied uniformly. low_mem_load (default off) and merge_loras (default on) - low_mem_load trades some speed for a lighter memory footprint when loading, and merge_loras controls whether stacked LoRAs get merged into one combined patch versus applied as separate layers; leave merge_loras on unless you have a specific reason to keep them separate.

Output is a single lora (WANVIDLORA), which feeds into your Wan model loader's LoRA input, or into the next chained WanVideo Lora Select By Name instance.

How to install it

Via ComfyUI Manager, search WanVideoWrapper and install, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper

then restart. The LoRA files themselves are separate downloads - HuggingFace is the standard source (kijai's own WanVideo_comfy repo hosts many of the speed and utility LoRAs the community uses) - placed in ComfyUI/models/loras.

Common issues & troubleshooting

LoRA seems to have no effect. Given the typed filename field, double-check for an exact-match typo before touching strength - a mistyped name is the single most likely reason a LoRA silently fails to apply.

Motion and detail got noticeably worse after adding a speed LoRA. That's the documented, real tradeoff, not a misconfiguration - plastic skin and flattened motion are what the community warned about running these at full strength on both passes. Drop to the low-noise-pass-only pattern at reduced strength before concluding the LoRA itself is bad.

Multiple LoRAs stacked and results look muddy or contradictory. Try merge_loras off to see whether the layered-application path behaves differently, and check blocks isn't set way too restrictively on one of the chained LoRAs if you added block editing anywhere in the chain.