Preview Selective Focus: Render the Lens or Blur a Depth Map?
Preview Selective Focus: Render the Lens or Blur a Depth Map?
A treatment board asks for three frames. Frame one: an opaque cutout in the left third of the shot, crisp, its edge crossing the bottle's shoulder. Frame two: the middle of the rack. Frame three: the label crisp, the cutout melted to a soft halo.
Two routes can produce that board. One assigns blur to pixels through a depth map. The other retains a scene, points a modeled camera at it, and lets an aperture decide what is soft. Both can yield a convincing sequence of stills, and the conviction is the problem. They are not two versions of one instrument. They know different things, they inherit different guesses, and they stop answering the question at different frames.
The choice between them is not about quality. It is about which frame in the pull you need to be able to defend, and what your source actually contains at that frame.
Identify what must become readable, and when
Three different questions get filed under "selective focus," and they have different tolerances.
Softness asks whether the background reads as out of focus. It is a look question and it is forgiving. A blurred plate is usually enough, because nobody will measure the bokeh against a lens that has not been chosen yet.
Readability asks when a specific detail crosses into legibility. The label's lower line, the emboss on the cap, the serial number. This is a threshold question, and threshold questions live in the middle of the pull, not at the ends. A still can look right at both ends and be wrong throughout the transition.
Reveal asks whether something new comes into view. This one deserves an early, honest check, because a focus change on a fixed camera does not do it. The set of rays that can reach the lens is fixed by where the surfaces are. A cutout standing between the camera and the bottle hides the same region at every focal setting; an open aperture only changes how the boundary between cutout and bottle is blended, and opening it moves the blend slightly outward, hiding more rather than less. The crispest state of the cutout is the state that hides the least. So if the treatment's dramatic beat depends on the bottle's mark becoming visible from behind the cutout, no render, map, or composite will produce it. Either the blocking changes or the beat does.
Once you know which of the three you are answering, inventory the source, because the source caps the preview before any setting is touched.
A flat plate holds one focus state, permanently. A plate plus separated layers lets you redraw the boundary between near and far by hand, which is a real gain in control and no gain at all in information — the layers still contain only what was photographed. A plate plus a depth map adds a per-pixel assignment of nearness, which is authored, estimated, or measured, and you should say which. A retained 3D scene holds geometry beyond any single view, so a camera can be re-aimed and a focus plane re-set without needing a photograph of the result. A measured scene — photogrammetry, a depth pass from a capture rig — is geometric but has its own holes and its own confidence.
The rule that follows is dull and stubborn: a preview cannot contain more than its source. Every method below is a way of spending what the source has.
For the rest of this piece, keep one construction in view. Nothing has been rendered, composited, or photographed for it; the geometry is stipulated so the comparison has fixed facts to work with. A matte light-gray cutout, roughly a leaf's silhouette, sits about 20 cm from a fixed camera. It fills the left third of frame, and its right edge reaches a little past frame center. A bottle with a printed label stands about 60 cm away, on a surface; the camera is aimed at the label, which sets the bottle just right of center with its upper body — the shoulder — along the cutout's right edge. Behind both, about 2 m out, a plane carries a repeating diagonal-stripe pattern. The camera is declared to be at f/2.8. The cutout's edge accordingly crosses two things: the bottle's upper body below, and the striped backdrop above it.
Three focal planes: the cutout for frame one, a plane between the cutout and the bottle for frame two, the label for frame three.
Build and qualify the depth-map route
Adobe's Camera Lens Blur takes an input map and a selected channel, and uses a Blur Focal Distance to name the map value that receives zero blur, with the rest of the map distributed around it. That is the whole mechanism, and the detail worth slowing down on is what the control is denominated in. The Blur Focal Distance is a value in the map. It is not a distance to the subject, and it does not become one because you typed the number you measured on set. Even a map calibrated to real scene distances — rare, and usually a deliberate choice — produces a preview that is a decision about a plate. "Focus on the label" becomes "put the label's map value at the focal distance," which is a statement about your map.
That matters more at the boundary than anywhere else, and our construction puts a boundary right through the middle of the shot.
A depth map carries one value per pixel. One value per pixel means one surface per line of sight. Where the cutout's edge meets the bottle, the map steps from a near value to a far value, and the position and sharpness of that step is something a person or an estimator decided. Blur, meanwhile, is a gather: each output pixel collects color from its neighborhood. At the step, the gather crosses it, mixing cutout into bottle and bottle into cutout across a band whose width follows the blur amount in force at that moment.
Two consequences are worth carrying into the frame review.
First, the mixing band is not a mistake, it is the approximation. A real defocused edge is also a mixture — some lens samples see the cutout, some see the bottle. The question is whether the mixture follows the aperture and the geometry, or follows a drawn contour. If the map's step sits three pixels off the true silhouette, you get a ribbon of wrongly-softened bottle next to a ribbon of prematurely-hard cutout. At low blur, nobody sees it. At the widest point of the pull, when the cutout is softest and its halo reaches furthest, that ribbon can be the width of a letter on the label.
Second, direction is not symmetric, and this pull runs one way. Blur can only move away from the sharpest state the plate contains, so a map serves the frames on the softening side of that state and never the frames that need sharpening. Focus the plate on the cutout — reasonably sharp at frame one — and a map can soften it into frames two and three, widening the halo as the declared focal plane recedes, and that transition is a plausible approximation. The same plate cannot give frame three its label: the bottle sits at the far end of the pull and the plate holds it soft, so the label's map value can be moved to the focal distance and the letters still will not resolve. Put the plate's sharpest state on the label instead, and frame one — the crisp cutout with the least hidden bottle — is asking for detail the plate never recorded. No map returns it. The soft cutout's edge is a contour you can find, roughly; the bottle light underneath the halo is gone.
An all-in-focus composite of the scene sidesteps the second problem neatly, and it is worth naming as the strongest flat source for this kind of pull, because both sharpness directions are then available for blurring. What it does not solve is the first problem, or the fundamental one: with a flat source, the split between cutout pixels and bottle pixels is drawn once, at the plate's focus state, and every subsequent blur can only mix across that drawn line. Notice, too, that an all-in-focus plate is itself a construction. If it was assembled from several captures, its boundary is already an authored decision that the depth map will then faithfully reproduce.
Compare a declared finite-aperture scene model
The other route does not soften pixels. It samples a point on the lens, traces a ray from that point through the pixel's position on the plane of focus, and asks what the ray hits. The thin-lens account in Physically Based Rendering (fourth edition, section 5.2.3) describes exactly this: a finite aperture, a focal distance, lens sampling, and depth of field that emerges from the geometry rather than being applied afterward. Blur varies with aperture, with focus setting, and with how far a surface sits from the focal plane.
For our construction, the operative difference is coverage. A ray that clips past the cutout's edge continues to the bottle. A ray that does not, stops. As the focal plane moves from the cutout toward the label, the width of the cutout's halo follows from the aperture and the declared positions, and the region of the bottle it covers is recomputed at every frame rather than gathered from a fixed plate. The stripes near the cutout's edge are computed the same way, sample by sample, so some rays see stripe and some see cutout and the blend between them is a consequence rather than a contour.
Two further properties of this route are easy to state and easy to overstate.
The scene retains detail at every focal plane. Because the bottle's geometry and label texture exist independently of any captured focus state, a frame focused on the label is genuinely resolved there. That is the real advantage for a focus pull, and it is the capability the depth-map route cannot borrow from any setting of its sliders.
And the model's authority runs only as far as its declared facts. If the cutout's distance from the lens is a guess, the render is an extremely precise picture of a guess — and the cutout's distance is the parameter that governs the halo width most aggressively, precisely because the cutout sits close to the lens, where distance from the focal plane changes fastest. More samples, a larger render, a longer render — none of that repairs a misplaced cutout. The render also inherits material assumptions: a stand-in bottle without a legible label texture cannot answer a readability question, no matter how many rays you cast.
Then there is the model itself. A thin lens is an idealization. It has no chromatic aberration, no field curvature, no diffraction, no sensor, no coating, no particular bokeh character. The PBRT section supplies a mechanism for understanding how a finite aperture produces depth of field; it is not a certification of any commercial lens, and a rendered frame is not optical evidence. It is evidence about geometry viewed through a declared aperture. Handing it over as "what the lens will do" is a category error with a very convincing surface.
Choose the route at the frame where it can fail
Both routes look fine at the ends. Choose by finding the frame in the pull where the treatment's most important detail sits closest to the widest halo, and ask each method to account for itself there. That is the frame where the two representations diverge, and it is usually somewhere in the middle.
Some workable rules, applied to our construction:
If the question is softness — does the background read as defocused, does the cutout sit in front — a flat plate and a depth map are enough, and the honest thing is to say so rather than spend a scene build on it. The boundary behavior near the cutout's edge is the only thing to inspect, and at moderate blur it will not be visible.
If the question is readability and the detail of interest sits comfortably clear of the boundary, the depth map is still the efficient route, provided the target frames are no sharper than the sharpest state of the plate. If the pull needs to arrive at a frame sharper than anything the plate holds, the flat route is being asked for information it does not have, and no setting will change that.
If the detail sits inside the mixing band at the widest point of the halo — a label line running within a few pixels of the cutout's edge is exactly this case — the boundary is the subject of the test, not a detail of it. Either compare both routes at that specific frame, or move to the scene render, and then check the cutout's declared position first, because that is the assumption the comparison rests on.
If the question is about reveal, no preview settles it. Check the geometry. If the detail is behind the cutout's crisp silhouette, it is behind it at every focal setting, everywhere, always.
If the question is what the actual lens will do on the day, neither route answers it. That is a bounded camera test: the real cutout, the real distance, the real aperture, the real label, and someone competent to read the result.
When you do compare the two, match the conditions that can be matched — same focal planes, same declared aperture, same framing, same output size — and then compare the transitions, not the end frames. Export the middle frames. Watch the order in which details cross into legibility, because the order is the thing the treatment is promising. And resist the pull of matched-looking images: two frames that resemble each other are not two representations of equal accuracy, and the resemblance is often strongest exactly where the two methods are quietly doing different things.
The two routes hide different assumptions, and that is the honest way to hold them. The depth-map preview inherits the plate's decisions — its focus state, its authored or estimated depth, its drawn boundary. The scene render inherits its own declared decisions — the cutout's position, the bottle's material, the aperture's idealization. Neither contains a capture. Both are arguments about one.
So close the comparison with three things written down: the route you are using, the frame where it is doing the least work, and the question it cannot answer. For a pull like this one, the route is usually determined by whether you need sharpness the plate lacks and whether the important detail crosses the boundary. The useful evidence is the boundary frame at maximum halo, examined for both the transition and the arrival. And the open question is the one that follows the preview out of the room: what the actual optic does at that distance, with that cutout, at that aperture — which stays unanswered until someone puts a camera there and looks.
Frequently asked questions
What are the three different questions filed under selective focus, and why do they have different tolerances?
Softness asks whether the background reads as out of focus; it is a forgiving look question. Readability asks when a specific detail crosses into legibility, a threshold question that lives in the middle of the pull. Reveal asks whether something new comes into view; a focus change on a fixed camera does not do it because the set of rays that can reach the lens is fixed by where the surfaces are. A cutout between camera and bottle hides the same region at every focal setting; opening the aperture moves the blend slightly outward, hiding more rather than less, and the crispest cutout hides the least.
Why can't a depth-map route make a bottle label sharper than the plate's sharpest state?
Blur can only move away from the sharpest state the plate contains, so a map serves the softening side of that state and never the frames that need sharpening. If the plate is focused on the cutout, the bottle sits at the far end and stays soft even if the label's map value is moved to the focal distance. An all-in-focus composite gives both sharpness directions for blurring, but the split between cutout and bottle pixels is still drawn once at the plate's focus state, and every later blur can only mix across that drawn line.
What does Blur Focal Distance actually control in Adobe Camera Lens Blur?
It names the map value that receives zero blur, and the rest of the map is distributed around it. It is a value in the map, not a distance to the subject, and typing a measured on-set distance does not change that. Even a map calibrated to real scene distances produces a preview that is a decision about a plate; focusing on the label becomes putting the label's map value at the focal distance, which is a statement about the map.
What is the main advantage and the main limitation of a declared finite-aperture scene model?
The route samples a point on the lens, traces a ray through the pixel's position on the plane of focus, and asks what the ray hits; the cutout's halo and the bottle region it covers are recomputed at every frame rather than gathered from a fixed plate. Because geometry and label texture exist independently of any captured focus state, a frame focused on the label is genuinely resolved there. Its authority runs only as far as its declared facts: a guessed cutout distance yields an extremely precise picture of a guess, and a thin lens is an idealization with no chromatic aberration, field curvature, diffraction, sensor, coating, or particular bokeh character. It is evidence about geometry viewed through a declared aperture, not optical evidence or certification of a commercial lens.
How should you choose between depth-map and scene-render routes?
Find the frame in the pull where the treatment's most important detail sits closest to the widest halo, and ask each method to account for itself there; that is usually somewhere in the middle. If the question is softness, a flat plate and depth map are enough. If it is readability and the detail sits comfortably clear of the boundary, the depth map is efficient provided target frames are no sharper than the plate's sharpest state. If the detail sits inside the mixing band, compare both routes at that frame or move to the scene render, and check the cutout's declared position first. If it is reveal, check geometry. If it is what the actual lens will do, neither route answers it; that needs a bounded camera test. When comparing, match conditions and compare transitions, not end frames.