AP Flow Composite

You've warped a frame with flow and computed an occlusion mask. Now you have two versions of every pixel - the original and the warped - and the composite node is the referee that decides, per pixel, which one shows. It's the last step of the pack's core temporal chain (APGetRAFTOpticalFlow → APFlowOcclusionMask → APApplyRAFTOpticalFlowMasked → here), and it's where a flow workflow either looks like clean video or falls apart into ghosting.

The idea is simple: warped pixels are what you want where the warp is trustworthy; original pixels are what you keep where it isn't. The art is deciding "trustworthy" - and this node gives you three levers to decide it.

The mechanism

APFlowComposite takes original_images, warped_images, valid_mask (from the apply node), and occlusion_mask (from the occlusion node), and builds an alpha per pixel. The base alpha is valid_mask × (1 − occlusion_mask) - trust the warp where the flow was consistent and not occluded. Then three dials refine that alpha:

• alpha_mode - flow_confidence (default), flow_confidence_x_mask, or mask_only. The first uses pure flow validity; the second multiplies by an optional effect_mask; mask_only makes the effect mask the whole story. For the README's recommended anti-blur setup you want flow_confidence_x_mask, with the apply node's warped_mask wired into effect_mask.
• use_difference_gate - the smartest feature here. When enabled, the node computes the per-pixel difference between warped and original, and only lets the alpha through where the difference exceeds difference_threshold (default 0.01). The point: if the warp didn't actually change a pixel, the composite shouldn't touch it either. This is what keeps background that shouldn't have moved from being repainted. The README recommends a low threshold like 0.005–0.02 for the anti-ghost setup.
• mask_threshold - binarizes the effect_mask above a value, for when you want a hard mask rather than a soft one.
• blend_strength, feather, invert_occlusion - blend amount (1.0 = full alpha, under 1.0 = partial), alpha edge feathering, and a polarity flip for the occlusion mask if yours is inverted relative to expectations.

Outputs

composited_images is your finished blend. composite_alpha is the alpha that produced it - a MASK you can preview or reuse, which is handy for debugging "why is my background getting touched" (answer: the alpha is too high there; check your difference gate).

The wiring that actually gets results

The README's "professional anti-blur compositing" is worth quoting because it's the difference between a soft blurry result and a crisp one:

1. Connect warped_mask from APApplyRAFTOpticalFlowMasked to this node's effect_mask.
2. Set alpha_mode = flow_confidence_x_mask.
3. Enable use_difference_gate = true with a low difference_threshold (0.005–0.02).

That combination confines the composite to genuinely-warped regions and leaves untouched background alone. Skip it and you'll get the classic symptom: background areas subtly blurring every frame because the composite is blending warped-but-identical pixels with the original at some alpha.

Installing it

Standard pack install: ComfyUI Manager (search "AP_OpticalFlow"), or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/adampolczynski/ComfyUI_AP_OpticalFlow
python -m pip install -r custom_nodes/ComfyUI_AP_OpticalFlow/requirements.txt

Restart. Dependency is torchvision>=0.15 only.

The one thing to check first

If your composite output looks like the warped frame is leaking in where it shouldn't, your occlusion mask polarity is probably flipped - try invert_occlusion. If it looks like nothing is being composited, your effect_mask is too tight or use_difference_gate is swallowing everything at a too-high difference_threshold. Both are quick dial-turns, not node-bugs.