🎭 LTX Likeness Anchor

The classic LTX identity failure: the face is right in the early frames, then drifts - subtly at first, unrecognizably by the end. LTXLikenessAnchor attacks that at the attention level, with per-block hooks on video self-attention (attn1) that pull generated tokens toward reference face features. It's the workhorse of TenStrip's pre-Reinforcer identity era, and it's still genuinely good - especially if you're on a merged/vanilla LTX2 checkpoint rather than the Best-Face-ID LoRA stack.

The matching uses centered cosine similarity (per-frame mean subtracted first, so common-mode features don't dominate), and v1.2 adds a directional pull mode that rotates token features toward the reference identity direction without changing magnitude - which is what preserves color and saturation.

The two reference modes

The design is deliberately flexible about where the identity features come from:

• latent_frame_0 (recommended for i2v) - uses frame 0 of the latent as the reference. In i2v workflows that's literally the conditioning start image, so you get identity anchoring with zero extra plumbing. frame_0_bbox lets you restrict the source to the face region.
• guide - reads the appended reference frame from a reference_info output, when you're running the Guide+Anchor combo.
• auto (default) - uses the guide if reference_info is wired, otherwise falls back to latent_frame_0.

So the minimal wiring is just: Model → LTXLikenessAnchor → Sampler. The strength (0.10, band 0.08-0.18 for directional pull; 0.30-0.60 usable for a hard identity lock) is the knob you'll actually touch.

The inputs that fix real problems

• pull_mode - directional (recommended) or additive (legacy). The additive mode pulls magnitude too and desaturates at high strength; directional is why the defaults changed.
• similarity_threshold (0.50) - how similar a token must be to get pulled. Raise it to narrow the effect to the strongest matches.
• decay_with_distance (0) - fade the pull strength from frame 0 to the last frame; 0.3-0.5 lets later frames breathe while early frames hold.
• late_block_falloff (0) - the useful one. The last 12 blocks (37-47) handle fine detail, and pulling still-image detail into video refinement steps over-sharpens. 0.3-0.6 disables or dampens late blocks, which is the documented fix for that overcrisp, texture-flattened look.
• skip_when_sigma_above (0) - disables the anchor on very noisy early steps.
• bypass - A/B compare without rewiring.

The README's honest warning: the author notes that with depth_curve=late_focus or very high strength, "variance-collapse side effects" are possible, and to test with clean state - hooks can leak between runs, so restart or bypass to get a true read.

The community reality check

If you're coming from TenStrip's face workflows, you'll have seen the identity-loss complaint pattern: small faces in wide shots lose identity, and the face-region conditioning can leave a visible box artifact. The fixes that recur in the threads are the same ones this node's parameters encode: crop the source closer to the subject (a full-body reference simply has too little face detail), use a second reference with a clean face read, and keep the upscale pass out of the identity path - several users report the upscaler is the fastest way to lose identity. This node is where you tune most of that.

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. For the newer Best-Face-ID setups the pack now points you at LTXFaceIdentityReinforcer instead - but for vanilla LTX2 identity anchoring, this is still the node.