O1 Blur is not defocus · Lab

1 The idea

A Gaussian blur turns bright points into soft fuzz, but a real lens turns out-of-focus bright points into crisp-edged discs (bokeh), especially strong lights like sparks.

Defocus needs linear values that aren’t clipped at 1, or the bokeh discs come out too faint. Click both nodes to compare.

2 Hands-on in the mini compositor

SkyVFX mini compositor · O1640×360 · float · 1001

3 Deep dive

What a lens does to one point of light

When a point is out of focus, its light does not land on one spot of the sensor. It spreads into a disc shaped like the aperture, called the circle of confusion. The disc is evenly bright inside and has a clear edge. That is bokeh.

A Gaussian blur behaves differently: brightest in the middle, fading out with no edge, so a light turns into haze. On flat areas the two look alike. The moment there are hard lights, like sparks, the difference is obvious.

Tool	Spread shape	Varies with depth	Use it for
Blur	Bell, no edge	No	Soft mattes, haze, diffusion
Defocus	Flat disc, clear edge	No (one size)	An element at one distance
ZDefocus	Disc, bladed, or your own image	Yes, from depth	CG or plates with depth
Convolve	Any image as the kernel	No	Real lens bokeh you photographed

Why it needs linear values that are not clipped

Defocus creates no new light; it only shares out the light that is there. Example: a one-pixel spark at 62, spread into a 14 px disc (about 154 pixels), leaves about 0.40 in each pixel of the disc. It still reads as a bright disc over the desert plate.

If the same spark was clipped to 1 first, say it came from a JPEG or went through a Clamp, each pixel gets only about 0.0065, nearly invisible. 62 times less light. That is why defocus belongs in scene-linear float, with the values above 1 still there.

1Linear EXR, values above 1 intact
2Defocus each layer
3Merge back
4Grain last

In our lab, Defocus averages 64 samples inside a disc. Same idea as Nuke’s Defocus, which doesn’t read depth either; for blur that follows depth, use ZDefocus

Where the disc size comes from, and which depth you have

Depth of field gets shallower with a lower f-number (wider aperture), a longer lens, or a closer focus. There is more depth of field behind the focus point than in front. Too shallow and the scene reads as a miniature. Take the lens and camera values from the matchmove instead of guessing.

Depth comes in two main kinds. Nuke’s ScanlineRender gives 1/Z (near = high, far → 0). Most renderers give real distance in scene units (near = small, far = large). Tell ZDefocus the wrong kind and you get almost no blur, or blur in the wrong places.

Depth type	Near	Far	ZDefocus math
1/Z (ScanlineRender)	high	towards 0	far = 0
Distance (renderer)	small	large	depth

Always sample the nearest and farthest thing first; you’ll know at once which kind it is

The edge between foreground and background

A CG robot stands sharp in front of a defocused desert plate. Run one ZDefocus over the merged image and the robot’s colours get smeared into the blurred background, leaving a ghost outline. A flat 2D image has no idea what is behind the robot.

The fix: split into depth layers, defocus each one with the focus linked to one master value, and merge back to front, or use deep if you have it. Know that even blur(FG) over blur(BG) is not physically exact where both layers are soft. It is an approximation people use because it is good enough in most shots.

Depth on an anti-aliased edge is an average of near and far, so it points at a distance where nothing exists. That shows up as a sharp outline inside something that should be soft. Extend the near depth a little past the edge, and do it exactly once.

1Split FG / BG
2Edge-extend depth once
3ZDefocus per layer
4Merge back to front

4 In real work

In Nuke

Defocus: one disc size for the whole image. Good for an element at one distance, like a burst of sparks all about the same distance from camera.
ZDefocus: set math to match the depth (1/Z = far = 0, real distance = depth), then switch output to focal plane setup to see the in-focus band.
ZDefocus filter type: disc, bladed or image. For a specific bokeh shape feed your own kernel image, or use Convolve on an element at one distance.
Turn depth of field off in the render and do it in comp, so you keep control and never blur twice.
Defocus the degrained plate, then add grain once at the end of the comp (why grain goes last: O5).

Common mistakes → the fix

✗ Using Blur for out-of-focus lights✓ Use Defocus or ZDefocus. Lights must become discs with an edge, not haze.
✗ Defocusing an element already clipped at 1 (a JPEG, or after a Clamp)✓ Use a float EXR with the values above 1 intact. If all you have is clipped, push the hot cores back up before the defocus.
✗ ZDefocus does almost nothing because math doesn’t match the depth✓ Sample near and far first, then pick far = 0 for 1/Z or depth for real distance.
✗ Grading the depth channel so it “looks nice”✓ Depth is distance, not a picture. Never adjust it by eye, or every defocus calculation goes wrong.
✗ One ZDefocus over the already-merged robot and plate✓ Split by depth, defocus separately, merge back to front.

Before you send the shot

Out-of-focus lights are discs with an edge, not haze
Defocus runs in linear float with values above 1 intact
You know whether depth is 1/Z or distance, and math matches
No sharp outlines or ghost edges around soft foreground objects
Depth of field fits the lens and doesn’t read as a miniature
Defocused areas still carry grain as sharp as the rest

5 Check yourself

One question to check yourself

Why must defocus happen before clipping at 1?

What you learned

🔒 Answer the question above and this card opens