LTX Feta Enhance (Ray)

LTX Feta Enhance (Ray) is a transformer-level enhancer that sharpens LTX output by boosting high-frequency detail during sampling - the "Feta" being short for a high-frequency enhancement scheme applied at specific layers. If you know LTX's reputation, you know the drill: fast, but with a tendency toward soft, smeary detail compared to Wan. Feta Enhance is one of the counterweights to that, and this node brings it into the raylight distributed pipeline.

Mechanically, it's a patch on the transformer. The node writes a feta_weight into the model's transformer options and tells the pack which layers to apply it to - the weight gets multiplied into the attention/feed-forward path at those layers, injecting extra high-frequency signal where it counts. The block_indices input controls the scope: empty (default) means all layers get the treatment; a comma-separated list like 8, 14 restricts it to specific blocks. The "all layers" default is deliberate - unlike the STG patch, Feta wants to run broadly.

The inputs

• feta_weight - float, default 4.0, and it goes negative. This is the strength of the enhancement. Positive sharpens; negative dampens high-frequency content, which is a genuinely useful direction if a workflow is overcrisp or noisy.
• block_indices - optional, default empty = all layers. Restrict to a subset when you want the enhancement targeted (for example, only the deeper blocks where detail forms).

Both feed through ray_actors and come back as ray_actors; the node sits in the model chain before your Ray sampler, and it composes with the other LTXV patches in the pack - the code merges with existing transformer options rather than replacing them.

Practical take

The default 4.0 is the author's chosen starting point, and it's a good one - start there, then bias-correct. If output still looks soft after Feta, bump the weight a couple points before you start stacking more enhancers; if you get grain or ringing artifacts, dial it back or drop it to 0 to confirm the patch was the cause. The negative range is the underrated part: with LTX's smudgy-motion reputation, a mild negative weight on a per-clip basis is a legitimate way to calm an over-grainy sequence.

One honest note: Feta is a detail enhancer, not a model upgrade. It makes what's already there crisper; it won't fix weak composition or prompt drift, which remain LTX's real weak spots (the community consensus is that seed-hunting beats prompt-wrangling for this model). Use it for what it's for - sharpness - and keep expectations in check elsewhere.

Install: ComfyUI Manager → "raylight", or clone + pip install -r requirements.txt. No model downloads.