LTXV Enhance A Video KJ

LTX-Video is Lightricks' famously fast video model, the one that runs near real-time on a 4090 when everything's tuned. Speed like that usually comes at the cost of detail and temporal richness. Enhance-A-Video is a technique to buy some of that back, and this node wires it into an LTX workflow with a single dial.

It's from kijai's KJNodes, and it's a straight port of the Enhance-A-Video research (the node description points at NUS-HPC-AI-Lab/Enhance-A-Video). The whole idea of Enhance-A-Video is that it's a training-free tweak to the model's attention that improves temporal consistency and detail, so you get a nicer clip without retraining anything or adding a heavy pass.

How it works

It patches the model's attention to strengthen the cross-frame relationships that fast sampling tends to wash out, then hands the patched model back. You slot it between your model loader and your sampler: model in, enhanced model out, and everything downstream runs as normal. There's a latent input too, because the enhancement is applied in the context of what's being generated.

The effect is dialed by a single weight. That's the appeal, one knob, no retraining, no extra model to download.

The inputs and outputs that matter

• model - the LTX model to enhance. The patched model comes back out.
• latent - the latent being generated, connected in.
• weight (default 4.0, range 0–100) - the strength of the enhancement. This is the only real tuning parameter. The default of 4 is a reasonable starting point; turn it up for a stronger effect, but back off if you start seeing artifacts or the motion going strange. High values push hard, and "more" is not automatically "better," find the point where detail improves before it starts fighting the model.

The output is model, the enhanced model, wired straight into your KSampler.

Installing it

Almost certainly you already have KJNodes if you're doing LTX work. If not: ComfyUI Manager → search KJNodes for ComfyUI → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-KJNodes
pip install -r ComfyUI-KJNodes/requirements.txt

Restart ComfyUI. No extra model download for this one, the enhancement is a code-level attention patch, not a separate weight file. You do of course need a working LTX-Video setup for it to attach to.

Common issues

No visible difference. Your weight may be too low, or the improvement is subtle on your particular clip. Try bumping the weight meaningfully (say from 4 toward 6–8) to see whether it's doing anything at all, then dial back to taste.

Artifacts or weird motion at high weight. Pull the weight down. Like most "boost" tricks, over-driving it introduces its own problems; the enhancement fights the model past a certain point. Treat the default as a floor and increase cautiously.

It's an LTX-specific node. This lives in the KJNodes/ltxv category and is built for the LTX-Video pipeline. Don't expect to bolt it onto an unrelated model and get the intended result, the patch targets how LTX's attention behaves.

Ordering. Make sure the enhanced model output feeds your sampler. If your KSampler is still wired to the raw model loader, the patch is doing nothing.