NABLA Attention KJ

KJNodes/experimental

NABLA_AttentionKJ

Experimental node for patching attention mode to use NABLA sparse attention for video models, currently only works with Kadinsky5

Example

JSON Example

{
  "class_type": "NABLA_AttentionKJ",
  "inputs": {
    "model": [
      "node_id",
      0
    ],
    "latent": [
      "node_id",
      0
    ],
    "window_time": 11,
    "window_width": 3,
    "window_height": 3,
    "sparsity": 0.9,
    "torch_compile": true
  }
}

This example shows required inputs only. Connection values like ["node_id", 0] should reference actual node IDs from your workflow.

Inputs

Name	Type	Status	Constraints	Default
model	MODEL	required	-	-
latent?	LATENT	required	-	-
window_time?	INT	required	min: 1	11
window_width?	INT	required	min: 1	3
window_height?	INT	required	min: 1	3
sparsity	FLOAT	required	min: 0, max: 1, step: 0.01	0.9
torch_compile?	BOOLEAN	required	-	true

Outputs

Index	Name	Type	Is List	Connection Reference
0	MODEL	MODEL	No	["{node_id}", 0]

How to connect to these outputs

To connect another node's input to an output from this node, use the connection reference format:

["node_id", output_index]

Where node_id is the ID of this NABLA_AttentionKJ node in your workflow, and output_index is the index from the table above.

Example

If this node has ID "5" in your workflow:

• MODEL (MODEL): ["5", 0]

Package: comfyui-kjnodes