Focus Pull

The rack focus is one of the easiest ways to make a clip feel like cinema instead of a generated GIF: open soft and let it resolve, or open sharp and let it fall away. Focus Pull, from the symbiotica-ai/comfyui-nodes pack, does exactly that on a VIDEO, and the timing is handled for you - which matters more than you'd think.

The mechanism is honest about what it is. This is not a depth-aware rack between two planes inside one shot. There's no depth model here. The whole frame moves together, a defocus that goes from "the lens is on this shot" to "the lens is off it." For the shots this node exists for, that's the right tool anyway: the subject already fills the frame, and the pull is the reveal. If you need a real rack between a foreground face and a background street, you want a depth-based node, not this one.

How it works

ffmpeg decodes the clip, each frame gets a disc convolution at a radius that follows a focus curve, and it's re-encoded. Your audio rides through untouched. The key detail is in the seconds input: the rack starts on the first frame going into focus and lands on the last frame coming out of it - there is nothing to line up by hand. Ask for 1.2 seconds and you get a 1.2-second pull that fills the clip. Ask for longer than the clip and it simply uses the whole clip.

Two directions, both useful:

• into focus - opens soft and resolves. The reveal. Great for a clip that starts mid-action, or for hiding the rough first frames of a generated shot behind a soft open.
• out of focus - opens sharp and falls away, ending soft. The "and then the memory fades" move.

The inputs that matter

• direction - into focus or out of focus. Pick by which end of the shot you want the viewer to remember.
• strength - how far out of focus the soft end goes. 0.3 is a gentle breath; 1.0 puts a face past recognising; 0 passes the video through untouched. Default 0.6 is a solid middle.
• seconds - the length of the rack, 0.1–10. This is the one to tune per edit. Short and punchy (0.5–1s) reads as energy; long and slow (3s+) reads as mood.

Output is one video, same geometry and duration as the input.

Install

Part of the symbiotica-ai/comfyui-nodes pack - search "Symbiotica" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/symbiotica-ai/comfyui-nodes.git symbiotica
pip install -r symbiotica/requirements.txt

Restart, and make sure ffmpeg/ffprobe are on PATH (the pack looks in PATH, Homebrew, /usr/bin) - without them every video node in this pack fails before doing anything.

Gotchas

Everything is re-encoded, so a long clip takes a while - that's the price of doing the defocus per-frame. And because the defocus is a uniform blur, don't set a huge strength on a clip with important detail at the edges; the whole frame goes soft together, so there's no shallow-depth look to be had here. If the pull feels too weak, remember it's a global rack: a small strength on a wide shot is barely perceptible, so nudge it up. For the full "this doesn't look AI" treatment, stack it with the pack's Film Grain and Camera Shake - grain plus motion plus a focus breath covers a lot of generated-video tells.