LZ Anima Artist Options (Advanced)

If LZAnimaArtistNode is the knob, LZ Anima Artist Options is the entire control room. It's an options bundle node: one ANIMA_OPTS output that plugs into the advanced_options socket on the Pack/CrossAttn/combined nodes and unlocks block-level layer selection, sampling-progress windows, EMA smoothing, low-rank consensus, static capture, and the anchor-Q experiment. Nothing to generate on its own - it's pure configuration that rides along with the injection.

What you're actually tuning

The full list is long, so here's the map, roughly from "you'll touch this" to "rarely":

• start_block / end_block - which cross-attention layers get patched, by index (0 = first layer, −1 = last). Restricting injection to mid/high blocks is the single most effective lever for "the style is bleeding into the composition." A layer_filter string (0,3,5-10,-1) overrides both if you want a non-contiguous set.
• start_percent / end_percent - window the injection across sampling progress (0.0 → 1.0). Style early, let structure settle late, or vice versa.
• normalize_weights - whether the per-artist weights act as relative proportions (on) or independent strengths (off). Auto-disabled if you used ::weight syntax in the artist chain.
• artist_ema_alpha - cross-step EMA smoothing of the injected attention (interpolate fusion only). 0.3–0.5 is light, 0.8+ heavy; it removes flicker-style instability across steps at the cost of lag. Auto-resets each sampling run.
• lowrank_k - only for combine_mode: lowrank_avg. The low-rank projection dimension: k=1 forces all artists along a single consensus style direction, k=2–3 keeps the main directions (recommended), k≥N degenerates to plain output averaging.
• artist_static_capture + static_capture_k - accumulate the artist attention over the first K steps, then freeze it. K=6 is the recommended default; the tooltip calls K=1 the "v18 single-point cache."
• artist_anchor_q + anchor_seeds_count / anchor_user_blend / anchor_deep_layer_threshold - the experiment: fix the query hidden state to a precomputed anchor (from a fixed-seed forward pass, ~1s overhead on first use) instead of letting it follow the user seed. anchor_user_blend mixes between pure-anchor and pure-user x; anchor_deep_layer_threshold restricts anchoring to shallow layers. Decoupling style from seed noise.

The pack is upfront that some of these are experimental, and the code actively guards the combinations: static capture + concat_with_base is incompatible (static auto-ignored with a warning), anchor_q + static capture are mutually exclusive, and anchor_q + concat_with_base disables anchor_q.

How to use it

Add the Options node, set the few knobs you care about, wire its advanced_options into the injection node's socket. Leave everything at defaults and it's a no-op - the injection runs exactly as if you hadn't connected it.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/liz-ils/ComfyUI-LZNodes

restart → LZNodes/Anima. ComfyUI Manager: ComfyUI-LZNodes. No extra dependencies.

The honest take

This is the "we ran out of room on the main node" node. Ninety percent of users should never touch it - the defaults are sensible and the strength/combine/fusion trio on the injection node gets you 95% of the results. The remaining 5% (layer targeting, EMA, low-rank consensus) is real but requires patience and a willingness to grid-test. Where it genuinely shines: if you've dialed in a combination that works, the Options node is how you save that recipe as a reusable config. Just don't start here - start with the combined node and one artist, then graduate.

One thing the docs don't scream loud enough: because all of this rides on model patching, every sampling run pays whatever cost your options impose (the anchor pre-run, the EMA state, the accumulation steps). Fiddly options mean slower generations. Budget accordingly.