ComfyUI-LTX-Attention-Toolkit

LTXAttentionCompareRuns

Diff two captures (read live from the registry by store_handle) for one metric, ranked by |A - B| per (block, head). To compare a dumped .pt, load it into a handle first with Store Load.

G Raw/LTX/Profiler · 0 inputs · 10 parameters · 2 outputs

LTXAttentionGridViz

Compact description unavailable for this node.

G Raw/LTX/Profiler · 0 inputs · 12 parameters · 2 outputs

LTXAttentionHeadCandidates

Combine the zscore diff of several metrics into one composite score per (block, head), to shortlist heads for a targeted intervention (e.g. Head Freeze) instead of eyeballing several separate heatmaps. Reuses LTXAttentionCompareRuns's store-loading/extraction/zscore logic.

G Raw/LTX/Profiler · 0 inputs · 9 parameters · 3 outputs

LTXAttentionHeadFreeze

Compact description unavailable for this node.

G Raw/LTX/Profiler · 1 input · 6 parameters · 1 output

LTXAttentionKeyMap

Compact description unavailable for this node.

G Raw/LTX/Profiler · 0 inputs · 12 parameters · 1 output

LTXAttentionMetricsViz

Compact description unavailable for this node.

G Raw/LTX/Profiler · 2 outputs

LTXAttentionQueryMap

Compact description unavailable for this node.

G Raw/LTX/Profiler · 0 inputs · 13 parameters · 1 output

LTXAttentionStoreInspect

Compact description unavailable for this node.

G Raw/LTX/Profiler · 1 input · 0 parameters · 1 output

LTXAttentionTimestepEvolution

Compact description unavailable for this node.

G Raw/LTX/Profiler · 0 inputs · 8 parameters · 2 outputs

LTXAttentionZoneAnalysis

For each (block, head), compute the ratio: attention_mass_vers_zone / fraction_zone_globale ratio > 1 → the head focuses on this zone ratio ≈ 1 → uniform attention (no preference) ratio < 1 → the head avoids this zone The mask uses pixel coordinates (final image) and is automatically resized to latent space (/32 for LTX). Requires store_mode=full_fp16 (or hybrid, for that block) in LTXAttentionCaptureSetup.

G Raw/LTX/Profiler · 1 input · 11 parameters · 2 outputs

LTXLatentDims

Compact description unavailable for this node.

G Raw/LTX/Profiler · 1 input · 0 parameters · 3 outputs

LTXQKVMultiplier

Compact description unavailable for this node.

G Raw/LTX/Profiler · 1 input · 7 parameters · 1 output

LTXQKVStoreInspect

Compact description unavailable for this node.

G Raw/LTX/Profiler · 1 input · 0 parameters · 1 output

LTXQKVTransfer

Compact description unavailable for this node.

G Raw/LTX/Profiler · 1 input · 13 parameters · 1 output

LTXResetPatches

Manual escape hatch for orphaned layers: if an intervention/capture node is deleted or rewired out of the graph, its last-registered layer on the shared diffusion_model is never explicitly unregistered (nothing calls unregister_layer for a node_id that stops executing) and would otherwise linger -- still consuming RAM and still applying on future runs. This node clears every registered layer and restores the pristine forward outright.

G Raw/LTX/Profiler · 1 input · 0 parameters · 1 output

LTXRFForwardSampler

Compact description unavailable for this node.

G Raw/LTX/RFInversion · 2 outputs

LTXRFReverseSampler

Compact description unavailable for this node.

G Raw/LTX/RFInversion · 1 output

LTXStoreDump

Save the current AttentionStore and/or QKVStore to a single .pt file.

G Raw/LTX/Profiler · 0 inputs · 3 parameters · 1 output

LTXStoreLoad

Load an AttentionStore and/or QKVStore section from a .pt file saved by Store Dump.

G Raw/LTX/Profiler · 0 inputs · 3 parameters · 2 outputs