LTX Feta Enhance

This one's a freebie in the good sense. FETA is the LTX port of Enhance-A-Video, a training-free trick that gets a video diffusion model to capture more detail and motion just by nudging its attention - no fine-tuning, no extra model, no meaningful speed hit. You add the node, set one number, and the output comes out crisper and more temporally coherent. When it works, it's the cheapest quality bump in the pack.

How it works

Enhance-A-Video's insight is that video models under-use their own cross-frame attention - the mechanism that keeps frames consistent with each other. FETA boosts the strength of that temporal attention by a single factor, which encourages the model to reinforce detail and motion that would otherwise get washed out. The original technique was demonstrated across HunyuanVideo, CogVideo, and others and picked up fast in the community (Kijai wired it in almost immediately); this node applies the same idea to LTX. It patches the model and passes it on - so it stacks cleanly with your sampler, guider, and other model patches.

The inputs and output

• feta_weight (default 4) - the enhancement factor. This is the whole node. Higher pushes the effect harder; the default of 4 is a solid starting point.
• attn_override (optional) - an advanced hook to target specific attention layers rather than the default set. Leave it unwired unless you know exactly why you're touching it.

Output is the patched MODEL.

Installing it

• ComfyUI Manager - search LTXVideo, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo, then restart.

Nothing extra to download - it's a pure math patch on the model you've loaded.

Common issues

Cranking feta_weight made it worse. More is not always better. Push the factor too high and you get over-sharpening, weird contrast, or motion artifacts as the attention boost overcooks. Treat 4 as the anchor and move in small steps; back off the moment detail turns to crunch.

No visible difference. The effect is more obvious on some content than others - busy, detailed, motion-heavy clips show it most; a slow static shot may barely move. Compare the same seed with the node bypassed and enabled rather than eyeballing one render.

It's fighting your other enhancers. FETA, STG, and heavy decoder noise all push toward "more detail," and stacking all three at high strength is how you land in crunchy, over-processed territory. Add them one at a time and keep the one that's actually helping.