Load LoRA

Load LoRA is the smallest node in the ComfyUI-Ruyi pack and the one with the least to say for itself: pick a LoRA, set a strength, wire it in. But if you've ever watched a video model output look wrong because the previous run's LoRA leaked into it, the way this node works is worth knowing. It's a two-line node with a real design decision behind it.

What it does

It attaches a LoRA from your ComfyUI/models/loras folder to the Ruyi pipeline and sets how hard that LoRA pushes. Ruyi is a 7B image-to-video model, so this is the "give the video a style or a character" control - same idea as the LoRA culture that grew around image models, just applied to a video transformer.

How it actually works

Here's the trick: this node doesn't merge anything at load time. Look at what it does and it just records the LoRA's file path and strength into the ruyi_model object - the same object Load Model handed you, plus a note. The actual merging happens later, inside the sampler: right before the denoising loop it calls merge_lora on every recorded LoRA, and immediately after generating it calls unmerge_lora to peel them off again.

That matters for two reasons. First, you can stack multiple Load LoRA nodes and they apply in order, which is how you'd combine a style LoRA with a character one. Second, the unmerge-after step means one generation's LoRAs don't contaminate the next - a real failure mode in some other video packs, and one less thing to babysit here.

Inputs and output

• ruyi_model - from Load Model or the previous node in the chain.
• lora_name - a dropdown of every .safetensors file in ComfyUI/models/loras. It can be left as None, in which case the node just passes the model through.
• strength_model (default 1.0, range −100 to 100) - the influence. The author's README suggests 1.0–1.4 for best results based on experience, which tracks with image-model LoRA habits. Negative values are allowed if you ever want to suppress a feature instead of amplify it.

Output: ruyi_model, so you can keep the chain going (it stacks cleanly with TeaCache and Enhance-A-Video).

Installation

No separate install - the node ships in the pack:

# ComfyUI Manager: search "Ruyi", install ComfyUI-Ruyi
# or:
cd ComfyUI/custom_nodes
git clone https://github.com/IamCreateAI/Ruyi-Models.git
pip install -r Ruyi-Models/requirements.txt

Then just make sure your LoRA files live in ComfyUI/models/loras (standard ComfyUI location) and the dropdown will pick them up - no model download needed for this node itself.

Common issues

The honest gap: Ruyi never built the LoRA ecosystem image models have. The model launched in December 2024 and got crowded out by Wan within a couple of months, so dedicated Ruyi LoRAs are rare - the KB's lora-training and ComfyUI-ecosystem docs note that the community's training culture moved on to the newer bases. If you're hunting for "the Ruyi style pack everyone uses," there mostly isn't one.

What that means in practice: either you trained the LoRA yourself (and know it fits this architecture), or you're on a Ruyi workflow someone shared that references a LoRA you need to track down. Either way the node itself is reliable - the failure mode people actually hit is a wrong or missing file in models/loras, which shows up as the dropdown being empty or a file-not-found at generation time. Check the folder path, then check the strength; 1.0 is a fine place to start.