🎯 LTX Latent Anchor

Watch a long LTX sampling run and you'll see the same thing every time: the first handful of steps lock in composition and semantics, then the late steps quietly drift - faces subtly change, the scene's physical logic loosens, prompt alignment fades. LTXLatentAnchor is an inference-time regularizer aimed squarely at that late-step drift. It snapshots the model's own representation at a mid-sampling step (when conditioning alignment peaks) and then pulls subsequent computation back toward that cached state, block by block.

The mechanism is what makes it different from the reference-injection nodes in this pack. It's not adding tokens or external images - it hooks the output of transformer_blocks[i].attn1 (video self-attention) and adds a residual that nudges each token's features toward the cached anchor when they're similar enough. The matching uses centered cosine similarity: raw cosine is dominated by common-mode features every token shares (positional encodings, scaffold features), so the node subtracts the per-frame mean first. That leaves identity-specific deviations that actually discriminate.

The inputs that matter

• model - required. Returns the same model with per-block attention hooks attached.
• sigmas - connect this from your scheduler (BasicScheduler, or the STG Guider's schedule). It's the difference between "predictable cache timing" and a fallback that's explicitly described as less reliable.
• strength (0.10) - pull magnitude. 0.05-0.15 is the typical band. Too high over-damps motion and flattens frame-to-frame variation.
• cache_at_step (6) - which sampling step locks the anchor. Mid-sampling (3-9 on 13-step schedules) is the sweet spot: conditioning alignment has peaked, fine refinement hasn't started.
• similarity_threshold (0.50) - minimum centered-cosine similarity for a token to receive pull. Raise to 0.60-0.70 to limit pull to the most similar tokens only.
• decay_with_distance (0) - per-frame strength decay from frame 0. 0.3-0.5 lets later frames drift more while early frames stay anchored.

That's the simple view. Flip advanced_mode and you get the research knobs - cache_mode (schedule / live_extraction / manual_calls), forwards_per_step, cache_warmup, anchor_frame, depth_curve, block_index_filter. You'll almost never touch these, but the one that matters if you see weirdness: forwards_per_step must match your actual sampler - 1 for distilled CFG=1, 2 for standard CFG>1 (cond + uncond), 3+ for CFG+STG. Get it wrong and the cache fires at the wrong moment.

Where it fits

Chain it on the model path before the sampler: Model → LTXLatentAnchor → Sampler. It coexists with the reference-token and face-anchor nodes - they hook different intervention points and use different sentinel attributes, so they don't collide. It's the basic variant in a family; LTXLatentAnchorAware adds reference-image spatial weighting if you want the pull concentrated on high-energy regions.

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. If your LTX clips look great at step six and wander by step thirteen, this is the node that was built for that exact feeling.