CompositeMotionBrushWithoutModel

CompositeMotionBrushWithoutModel is the same merge as CompositeMotionBrush, minus the model input and minus the resulting headache of loading a 9.5 GB checkpoint just to combine two tensors. If you've ever squinted at a node that demands a model it never actually uses, you'll appreciate why this variant exists.

The behavior is identical: motion_brush_layer1 wins at every pixel where it has any motion (a vector larger than a rounding error), and everywhere it's still, motion_brush_layer0 shows through. The output is sized from layer0's shape - the node reads model_length, height, and width straight off the first brush - so it doesn't need a model to know what it's working with. Where CompositeMotionBrush can't even run without a loaded checkpoint, this one works on brushes alone.

Inputs

• motion_brush_layer0 - base layer, sets the output size.
• motion_brush_layer1 - overlay layer, overrides wherever it moves.
• mode - dropdown with one option, override. Same as the model-requiring twin.

Output

One merged MotionBrush, sized to layer0. Same downstream wiring: DragNUWA Run MotionBrush, or back into another composite for multi-layer stacks.

When this is the right node

Any time you're building a brush library or planning a composite before you're ready to pay for inference. Because the brush generators - InstantCameraMotionBrush, Load MotionBrush From Tracking Points Without Model, the optical-flow "Without Model" loader - also don't need a checkpoint, you can design an entire multi-layer motion field with zero model load, preview the layering, and only spin up Load CheckPoint DragNUWA + DragNUWA Run MotionBrush for the final run. That's a genuinely better workflow than the model-gated version, and it's the one I'd default to.

The trade-offs

Same resolution and frame-count discipline as the model version - mismatched brush shapes give you garbage merges, and since this node sizes itself off layer0, a wrong-sized layer1 either bleeds or vanishes. And same SVD-era reality on the output end: the compositing is precise, but what the model does with your carefully layered flow field is still 2024-era video generation, so expect softness and occasional warping on ambitious moves.

If you're new to the pack, start here and only reach for CompositeMotionBrush when a workflow you've loaded explicitly wants the model-typed version - the "Without Model" naming is the pack's quiet way of admitting which one it should have shipped first.