Enhance-A-Video for Ruyi

If TeaCache is Ruyi's speed plugin, this is the quality plugin. "Enhance-A-Video for Ruyi" wraps a training-free technique from NUS that boosts texture and detail in generated video - think of it as the video-model equivalent of a sharpening pass you can switch on mid-generation instead of after. It won't fix a weak start image, but it makes good Ruyi output look less soft.

What it is

Enhance-A-Video is a plug-in module (from NUS-HPC-AI-Lab) that improves video fidelity without any fine-tuning - no extra weights, no model download, just changes to how sampling runs. The Ruyi pack integrates it as a RUYI_MODEL pass-through node, the same pattern as TeaCache: drop it into the chain before the sampler, and the sampler activates it during the denoising loop.

How it works

During sampling it perturbs and re-blends the feature maps flowing through the transformer so that texture detail comes through more strongly in the final frames. You don't need to understand the math to use it, but you do need to respect the two knobs that control it:

• weight (default 1.0, range 0–100) - how hard the enhancement pushes. The tooltip is blunt about this: it should be smaller than 5 or 10 to get better results. This is the one that gets people into trouble.
• skip_start_steps and skip_end_steps (both default 0) - leave the first or last n steps untouched. Unlike TeaCache, they default to off, so the enhancement applies across the whole run out of the box.

Plus enable (default on), which is your kill switch.

Inputs and output

Inputs: ruyi_model, enable, weight, skip_start_steps, skip_end_steps. Output: ruyi_model, chainable.

Same wiring rule as TeaCache: this node only records settings; the sampler does the actual work, and it only runs if the node is in the chain. Remove it and Ruyi quietly samples without the enhancement. It also plays fine alongside TeaCache - the sampler initializes whichever plugins are present in the ruyi_model it receives, so Ruyi_LoadModel → Ruyi_TeaCache → Ruyi_EnhanceAVideo → Ruyi_I2VSampler is a perfectly valid setup: faster and sharper.

Installation

Ships with the pack, so there's nothing extra to install:

# 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

Add the node between Load Model (or any other ruyi_model producer) and the sampler.

Where people get burned

The trap is the weight slider. Because the range goes to 100, it looks like more is better - and then your video comes out overcooked: crunchy textures, waxy skin, detail that looks processed rather than enhanced. The author's guidance is under 5 or 10 for a reason; most people never need more than 1–3. Start at the default of 1.0, and if the output still looks flat, nudge it up a half-point at a time rather than jumping to double digits.

The other thing to know is what it isn't: it's not a fix for a blurry or low-quality input frame. Garbage in, slightly sharper garbage out. If your base image is soft, upscale or re-roll it before you blame the enhancement - and keep your expectations in line with a model that, by mid-2026, has been outclassed by Wan and LTX for raw quality. What this node gives you is a genuine detail bump on the model you already have, for free, no retraining required.