🎯 LTX Latent Anchor Aware

Plain LTXLatentAnchor pulls every sufficiently-similar token toward the cached mid-sample state, with the same strength everywhere. That works, but it's not smart about where the pull matters. A sky doesn't need anchoring - an edge, a face, a piece of detail does. LTXLatentAnchorAware is the upgraded variant: it takes an optional reference image, computes a spatial energy map from it (edges, faces, fine detail = high energy), and weights the anchor pull so high-energy regions get held tightly while flat regions drift freely.

That's a meaningful quality difference in practice. Anchoring a whole scene uniformly can flatten motion and dull texture - the "everything is a bit too stable" look. Anchoring only where detail lives keeps the composition stable without putting a velvet glove on the whole frame.

What changed vs the basic Anchor

Same core mechanism: snapshot the model's representation at cache_at_step, pull later computation toward it via centered-cosine-matched residuals on attn1 output. The addition is the energy channel:

• reference_image + vae - wire a reference image (ideally the i2v conditioning image or a composition reference) and the LTX2 VAE. The node encodes it and derives the spatial energy map.
• energy_latent - alternative to image+VAE: hand it a pre-encoded LATENT to skip re-encoding if you already have one.
• energy_threshold (0.30) - the gating cutoff. 0 = uniform energy (behaves like plain Anchor). 0.30 = regions above the 30th percentile of energy get pull. 0.80 = only the top 20% regions. This is your main dial for how "aware" the anchoring is.

Everything else is shared with the basic variant: model in, MODEL out, sigmas strongly recommended for predictable cache timing, strength (0.10, band 0.05-0.20), cache_at_step (6), similarity_threshold (0.50), decay_with_distance, bypass, debug, and the advanced_mode research knobs (cache_mode, forwards_per_step, cache_warmup, anchor_frame, depth_curve, block_index_filter).

Wiring and caveats

Model → LTXLatentAnchorAware → Sampler
              ↑
     reference_image + vae  (or energy_latent)

The one gotcha that's easy to trip over: vae is only required when reference_image is wired, and the energy_threshold note is worth internalizing - 0.50 means above-median energy only, so for a busy scene you may want 0.30-0.40 to avoid anchoring into a checkerboard of tiny patches. If you leave the reference inputs empty entirely, the node just behaves like the basic Anchor with extra parameters doing nothing.

This is the variant the README recommends over the basic one for most cases - it's the same tool with a genuinely useful refinement, not a different class of node. For faces specifically, pair it with the face-targeted identity nodes (they hook the same attn1 layer via different sentinel attributes, so they compose cleanly). Install is the pack-wide clone (10S_Nodes into custom_nodes, restart, or ComfyUI Manager → "10S-Comfy-nodes"), no extra deps, LTX2-class models only.