// Imaging Fundamentals

Continuous Focus, Explained: How Your Microscope Stays Sharp For Hours

Getting into focus once and staying in focus for hours are two different problems. Most of us only notice the difference the hard way, after a long run, when a beautiful dataset turns out soft at the exact moment we cared about most.

By Jacob Gewerth|July 20, 2026|9 min read

01Focus drift: the quiet failure

On any real microscope, focus drifts. The causes are boring and mechanical: thermal expansion as the room and electronics warm up, slow settling of the stage or nosepiece, tiny movements in mounting hardware, changes in immersion media.

Each of those shifts the distance between objective and sample by a few hundred nanometers at a time. Over 30, 60, or 90 minutes those small nudges add up. The field that started perfectly sharp at time zero is visibly soft by the end of the run, even though you never touched the focus controls.

You often do not see it happening. You see the before and after. Frame 1 looks great, frame 600 does not, and you cannot re-run the experiment.

02What continuous focus actually is

Continuous focus is a second focus system running in parallel with your imaging. Instead of looking at the image to decide whether focus is good, it watches a reflection from the sample carrier, usually a coverslip or glass-bottom well, using a separate light path.

A common design sends a near-infrared beam through the objective, down to the coverslip or vessel surface. The beam reflects off that interface and comes back up to a small detector, which measures exactly where the reflection lands. The imaging camera never sees this beam, so the system can run all the time without adding exposure or bleaching.

The key idea is simple. Instead of asking does this picture look sharp, the hardware asks did the coverslip move relative to the objective, hundreds of times per second.

03The lock loop, step by step

Most reflection-based continuous focus systems run a tight feedback loop:

1. Watch where the reflection lands right now.
2. Compare that position to a stored lock target.
3. Correct by nudging the objective in Z to close any gap.
4. Repeat continuously in the background.

That is what closed loop means here. Measure, correct, measure again, without stopping. Because the system is looking at a bright reflection on a dedicated detector, it can run very quickly, often on millisecond timescales, long before drift becomes visible in the image.

04Locking to glass vs choosing the sample plane

There is an important catch. The hardware does not lock onto your cells. It locks onto glass. Your actual sample of interest usually sits some distance above that coverslip or plate bottom.

To deal with that, continuous focus systems expose an offset: a user-chosen Z distance between the lock point on glass and the imaging plane in the sample. You focus by eye or with software autofocus, turn the hardware lock on so it grabs the coverslip, then adjust the offset until the locked position lines up with the plane you care about.

Once that offset is set, every return trip to that position should land on the same imaging plane, assuming the coverslip and sample geometry stay the same. The hardware is very good at holding this plane. It cannot tell you whether you chose the right one biologically. That still depends on your setup and on how you defined the offset.

05Where continuous focus struggles

To keep that lock, the system needs a strong, consistent reflection from the sample carrier. Glass-bottom dishes and coverslips are the easiest cases. Some plastics give weaker or inconsistent reflections, and the signal can be lost entirely if the interface is rough, tilted, or contaminated.

Other common failure modes:

• Wrong coverslip thickness, which weakens or shifts the reflection enough that the detector no longer tracks it reliably.
• Thick media or tall samples that push the desired imaging plane outside the hardware working Z range, even when the reflection is still visible.
• Multiwell plates where the system chases the well edge instead of the cells, especially near walls and menisci. A single global offset is often not enough. Per-well or interpolated offsets give better results.
• Immersion objectives where the oil or other medium pulls away during a long run, breaking the optical path and dropping the lock.

06Hardware vs software autofocus

Both aim at the same goal, sharp focus. They get there in different ways, with different trade-offs.

Image-based (software) autofocus

• Looks directly at the image, typically by scanning a small Z stack and scoring each plane with a sharpness or contrast metric.
• Costs exposures, which can add bleaching or phototoxicity for sensitive samples.
• Takes hundreds of milliseconds to seconds, depending on plane count and field of view.
• Usually runs only when asked. During setup, between stage moves, or at defined intervals.

Reflection-based (hardware) continuous focus

• Looks at a reflection from the coverslip or vessel surface, not at the image.
• Adds no imaging exposures, so it can run continuously without extra bleaching.
• Is much faster, often updating on millisecond timescales, in the background while you image.
• Is excellent at holding a defined plane, but depends on correct setup and on the quality of the reflective interface.

In many modern systems, image-based autofocus is used during setup or at sparse reference points, while hardware continuous focus maintains focus in between. In plate and slide workflows, using the hardware correctly, with per-well or per-region offsets and sensible limits, often removes the need for frequent image-based checks during a run.

07Using continuous focus safely

Getting continuous focus dialed in is mostly about a sensible start and clear guardrails:

1. Focus the sample by eye or with software autofocus at a representative position.
2. Turn the hardware lock on so it grabs the coverslip or vessel surface and stabilizes the reference.
3. Adjust the offset so the locked position corresponds to your desired imaging plane.
4. Watch the lock indicator during your first runs. Note how far the system has to move and how often it drops lock.

Because the lock reports its own state, it can act as a safety fence. If it constantly pushes to the edge of its working range or keeps dropping out entirely, the problem is not just drift. It is a sign that something in the mounting, immersion, or plate geometry needs attention.

08When continuous focus makes a difference

It mainly changes your life in workflows where drift would otherwise quietly destroy data:

• Long timelapse imaging, where hour six should look like hour one.
• Multipoint and multiwell imaging, where you revisit many positions and expect them to return to the same plane.
• Large tiled scans and mosaics, where a few soft tiles in one corner ruin an otherwise beautiful stitch.
• Live-cell and light-sensitive work, where you want to avoid any extra focus sweeps that add dose.
• Long unattended runs, where nobody is there to catch a slow drift in real time.

For one-off snapshots or short sessions where you can refocus by hand or with a quick autofocus sweep, continuous focus is less critical.

We build Nebula so this kind of thing stays understandable. Hardware focus, software autofocus, and interpolated offsets all live in the same acquisition workflow, on whatever hardware you already own, instead of being buried in four different vendor dialogs.

Interpolated focus, explained simply

How the software keeps a whole part sharp from a few measured points, so no defect slips through soft across a large scan. No focusing at every tile. In plain language.

01 What it is

Interpolated focus lets the software keep a large part sharp without focusing at every tile. You measure true focus at a handful of spots. It fits a surface through those points and reads off the right height everywhere in between, so every feature you are inspecting stays in focus edge to edge.

Think of it as a survey. Measure the height at a few spots, draw the terrain between them, and you can predict the height anywhere you did not measure. Here the terrain is the surface of your part, and the survey stakes are the few places you focused by hand or by autofocus.

Once the surface is set, every tile in the scan gets its own predicted height. No single compromise height that only works in the center. The far corner of the part gets the height the far corner actually needs.

The alternative is to run autofocus at every tile. That works, but it is slow, and slow is expensive when you are inspecting a large part or a full lot. A surface pays the focusing cost once, up front, then coasts across the whole scan.

02 Why one focus height fails

A part looks flat. It rarely is. It sits at a slight angle in the fixture or on the stage. It has its own warp, machining marks, coating thickness, and step heights, whether that is solder and traces on a board, a plated surface, or the relief on a fracture face. On a failure sample, the surface you care about is the least flat of all.

Set one focus height and inspect a small field, and you will never notice. Scan a large part and the edges tell on you. The plane you focused on in the center has drifted away by the time you reach the far corner. Those tiles come back soft, and a soft tile is a missed crack, pore, or particle.

The tilt does not have to be large. A slope of a single degree is enough to pull the edge of a big scan out of focus. The bigger the part, the worse it gets.

The fix is to stop describing focus as one number and describe it as a surface instead. Then every tile gets its own height, and the whole part stays sharp edge to edge.

03 Placing the points

This is the part you control. You decide how many points to measure and where to put them.

Match the count to the shape of your part. A flat part with a slight tilt needs only a few points. A plane through three or four spots describes it fine. A warped, stepped, or curved part needs a denser grid, because a plane cannot follow a curve.

More points is not always better. Past a certain density, extra points stop describing real shape and start chasing noise in each measurement. Use enough to capture the shape, not so many that you overfit.

Where you put them matters as much as how many. Anchor the corners first, so the fit is pinned at the edges. Then add points over the features you are inspecting, the weld, the die, the fracture, the connector, and anywhere the height changes fast, like a step or an edge. Keep the area you care about inside the ring of points. The surface is only trustworthy between the points you measured. Past the outer edge it is a guess.

04 The heat map

Once the surface is fit, the software colors it by height. That heat map is the fastest read you have on the shape of your part and whether to trust the fit.

Even color across the whole map means a level part. A single height would nearly have done the job.

A smooth gradient from one side to the other is a tilt. The part is not sitting square in the fixture. Normal, and exactly what the surface is there to correct.

A bullseye or dome means the part is warped or bowed. Common in thin parts, boards, and coated surfaces. A plane cannot describe that, so switch to a curved fit.

A speckled, noisy map means your points disagree with each other. Look for a mis-focused point or a specular hot spot, fix it, then add a few points to confirm.

The colors show relative height, not absolute thickness. The mid tone is your reference plane. Warmer is higher, cooler is lower. Read that way, the same map also tells you the part’s warp and coplanarity at a glance.

05 The math, lightly

You do not need the math to use this. But it helps to know what the software is doing.

Every method solves the same problem. Find a surface, z = f(x, y), that agrees with your measured points. They differ only in how much the surface is allowed to bend.

The simplest is a tilted plane. Three points define it. Solve for the slope in x, the slope in y, and the height:

Give it more points than it needs and there is no exact answer that hits them all. Least squares finds the surface closest to all of them at once. The payoff of the extra points is that one bad measurement no longer throws the whole fit. The surface averages through it, and helps you spot the outlier.

To follow curves, add terms. An x·y term lets the surface warp (bilinear). Squared terms let it bow (biquadratic). Higher orders follow more complex shapes (bicubic). The rule is simple: an n-order fit needs (n + 1)² points. A plane needs 3, bilinear 4, biquadratic 9, bicubic 16.

06 How to use it well

Short version.

Anchor the corners first, then add points over the features you inspect. Keep the area of interest inside the points.

Match point density to shape. A few for a tilt. A grid for a warp. Not so many that you chase noise.

Read the heat map before you run the scan, not after. Even color is fine. A bullseye means switch to a curved fit. A speckled map means fix a bad point. Catch a tilted fixture or a warped lot before you commit a batch to the scan.

Pair it with autofocus. The surface gives autofocus a close starting height at every tile, so the search is short and never wanders off looking for a peak on a reflective surface.

And do not trust the surface past the outer points. Interpolation between your points is solid. Extrapolation past them is a guess.

None of this needs a software background. It needs someone willing to lay out what the software is actually doing instead of leaving it as a black box. Keeping a whole part sharp for inspection is the software’s job, not yours to chase tile by tile.

This is how focus interpolation works in Nebula. You place the points, Nebula fits the surface and shows you the heat map, and the scan holds focus across the whole part so no defect slips through soft. It is hardware-agnostic, so it runs on the inspection scope you already have instead of locking you to one vendor’s stack.

Software autofocus, explained simply

How the software finds your sharpest plane using the picture itself. No sensor, no focus motor. In plain language.

01 What it is

Software autofocus lets the computer find focus on its own. No focus motor. No extra sensor. It judges the picture itself and keeps the sharpest plane.

The loop is simple. It moves the stage to a height, takes a frame, and rates how sharp that frame is. Then it does the same thing at the next height. Move, capture, score, repeat.

After it has tried several heights, it keeps the one with the best score. You do not pick the height. You give it a range, and it finds the sharpest plane inside that range on its own.

Hardware autofocus is the other kind. It uses a separate sensor that watches the coverslip and corrects in real time. It is fast and it holds focus well during a time-lapse. But it needs that extra sensor, and not every scope has one. This article is about the software kind, which works from the image you are already taking.

02 Measuring sharpness

Everything depends on one number. The focus score answers a single question. How sharp is this frame?

In focus, edges are crisp and neighboring pixels differ a lot. A dark pixel sits next to a bright one. Out of focus, everything blends together and those differences shrink. Sharp frames score high. Blurry frames score low.

There are three common ways to turn a frame into that number.

Contrast. Edge differences. The safe default, and it works for most samples.

Intensity. The brightest pixels. Good for a few bright beads on a dark field, where there are not many edges but there is a clear signal.

Variance. The spread of brightness across the frame.

Start with contrast. Switch to intensity when your sample is sparse and bright.

03 The focus curve

Score every height in your range and plot the scores against position. You get a focus curve. It climbs to a peak, then falls. The top of that hill is best focus.

Reading the curve tells you whether to trust the result. A real peak stands out from the flat parts around it. That means there is genuine structure in the field to lock onto.

A bad curve is jittery and flat. No clear peak, just noise wobbling around the same level. That usually means there is nothing in the field to focus on. An empty region. A blank patch of dish. A channel with no signal.

This is where software autofocus most often fails, and the fix is counterintuitive. When the curve is flat, do not move at all. Better to hold position than chase a peak that is not there. One clear peak is the goal. No peak means stay put.

04 Coarse then fine

Checking every micron over a wide range is slow. Forty microns at one-micron steps is forty frames every time autofocus runs. Multiply that across a tiled, multi-point, time-lapse run and you spend the session focusing instead of imaging.

So the search runs in two passes.

The first pass is coarse. Big steps over the full range to find the rough neighborhood of the peak.

The second pass is fine. Small steps in a narrow window around the winner to nail the exact height.

Coarse finds the neighborhood. Fine finds the exact address. Same answer as a brute-force scan, a fraction of the frames.

05 Every step costs light

Here is the part people forget, and it matters more than speed. Every step in the search is one more exposure. For fluorescence, every exposure is more bleaching. The routine is not free. It spends signal from the exact sample you are trying to preserve.

Put numbers on it. A coarse plus fine search might cost around 16 shots. A fine search at the same small step size across the whole range might cost 40 for the same answer. That is more than double the light, every time autofocus runs, with nothing to show for it but a dimmer sample.

Fewer shots means less light on the sample. That is the real reason two passes win. Not just faster. Gentler on the thing you care about.

06 How to set it up well

Short version.

Set a sensible search range. Wide enough to contain focus on a bad day, narrow enough that you are not scanning empty space.

Pick contrast unless your sample is sparse and bright, then use intensity.

Use coarse then fine, not one fine pass over the whole range. Same accuracy, far less light.

Trust a clear peak. Ignore a flat one. If the curve is noise, hold position.

And treat every focus frame as a frame off your light budget. Spend them like they cost something, because they do.

Pixel size, resolution, and field of view, explained simply

A few weeks ago I posted a one-page cheat sheet on machine vision basics. It did better than anything I have made, and the comments turned into a free education. Optical engineers and PhDs jumped in with the next layer: object-space pixel size, why magnification is not resolution, when to bin. I did not know all of it cold. So I went and learned it, and I built the follow-up. This is that follow-up, written the way I needed it explained.

01There are two pixel sizes

The first thing that tripped me up was the phrase pixel size. There are two of them and they are easy to mix up. Sensor pixel size is physical. It is the pitch of one pixel on the chip, often 6.5 microns on an sCMOS camera, and you read it off the spec sheet. Object-space pixel size is what one pixel covers out on your sample. That second number is the one that matters.

You get it by dividing the sensor pixel by your total magnification. The trap is the coupler. A 0.63x C-mount adapter on a 60x objective is really 37.8x, not 60x, and if you forget it, every number after that is wrong. If you only remember one thing from this whole post, remember object-space pixel size.

02Resolution is the lens, sampling is the camera

Resolution and sampling are not the same thing, and that took me too long to see. Optical resolution is what the lens can actually resolve, set by numerical aperture and the wavelength of light. A 1.4 NA lens in green light resolves down to about 0.18 microns, and the camera cannot beat that. Camera resolution is just how finely you sample that image with pixels. Your system is limited by whichever one is worse.

The match point is Nyquist: make your object pixel about half the optical limit. Sample coarser than that and two close things blur into one. That is undersampling. Sample much finer and you only make bigger files for no new detail. That is oversampling.

Both have their place. Undersampling is fine for screening, counting, tracking, and low-light work where you want speed and reach more than fine detail. Oversampling earns its cost for single-molecule localization, super-resolution, and deconvolution. Most everyday research lives at matched.

03Bigger pixels see more light

A pixel is a bucket for light. A bigger pixel has a wider mouth, so at the same exposure it collects more photons. More photons means a stronger signal against the same noise, which is a higher signal to noise ratio. Faint structure that a small pixel would lose in the noise starts to show up.

The catch is resolution. Spread the same sensor over bigger pixels and you sample coarser. Sensitivity goes up, fine detail goes down. Binning is how you make bigger pixels on demand. It electronically groups neighboring pixels and sums their signal into one larger superpixel. Bin when you need sensitivity and can give up resolution, when you want more speed, or to shrink file sizes. On older CCD cameras binning also cut noise. On modern sCMOS it helps file size and speed more than noise.

04Field of view is the trade for magnification

Field of view is how much of the sample you see at once. It is the sensor size divided by the total magnification, so magnification and field of view pull against each other. Higher magnification gives you finer sampling and a smaller window. Lower magnification gives you a bigger window and coarser sampling. There is no setting that gives you both.

For screening or tiling a slide, a bigger field means fewer tiles, faster runs, and less data to store. For resolving fine structure, you accept the smaller field. Pick the side your experiment needs and stop fighting the tradeoff.

05The sensor and the mount

The last piece is the camera mount, and it surprised me. A microscope projects a round image of a fixed size, called the image circle. Your sensor is a rectangle that has to fit inside that circle. Make the sensor too big and the corners fall outside the circle and go dark. That is vignetting, and it is why most scientific cameras are square or 4 by 3 rather than wide.

A reducing coupler, like a 0.5x or 0.63x, shrinks the projected image to fit a larger sensor, and as a bonus it widens your field of view. So sensor size, aspect ratio, and the coupler are really one decision: how to fit a rectangle inside a circle.

The imaging session basics I wish someone had just explained to me

I did not come from a research background. Army, then deputy sheriff, then software. So the first time I sat down at a microscope and opened the acquisition software, I froze. A wall of buttons, and none of them meant anything to me. Everyone around me moved like it was obvious. I spent that whole first session scared to touch anything in case I broke the experiment. This is the walkthrough I wish I had that day.

01The wall of settings is just one loop

This is what froze me on day one. I opened the software, counted something like forty settings, and assumed I had to understand all of them before I could press start.

Here is what nobody told me. Three things happen every time you image. Light hits your sample. The microscope captures what comes back. The software records it and saves it. That set of instructions is your acquisition protocol. Once I understood that every setting on the screen is just a way to control one of those three steps, the wall stopped being a wall. It was a control panel for a process I already understood.

02“MDA” is just five decisions

Someone told me to “just set up an MDA real quick” and walked away. I nodded like I knew what that meant. Multi-dimensional acquisition sounded like something you needed a credential to do.

It is five decisions. Which color of light you are using, your channel. Where on the slide you are imaging, the XY position. How deep you are going, the Z position. Whether you are imaging over time. And where your files land, the save location. Channels plus Z plus time captured at once is everything MDA has ever meant. You can start with one channel, one position, and one snapshot. That still counts as a session. Add the rest when you need it.

03Exposure is the one that will bite you

I blew out my first real sample and did not even know it. The bright parts looked solid white, so I kept fiddling with the display, trying to bring the detail back. It was already gone.

Exposure is just how long the camera stays open collecting light. Too short and your real signal sinks into the noise. Too long and the bright areas clip to pure white, and that detail never comes back. With live cells, long exposures also stress the sample. The habit that saved me is simple. Start low. Watch the live preview. Raise the exposure until the image looks right without going white. Get that right before you touch anything else.

04A Z-stack is not as advanced as it sounds

“Z-stack” sounded like something only the senior postdoc was allowed to run. I avoided it for weeks.

Here is the whole idea. If your sample has thickness, one image only catches one depth. A Z-stack grabs a slice at each depth and stacks them into a single 3D image. You set the top of your sample and the bottom. The software steps through everything in between on its own. Flat sample? Skip it entirely. That is all there is to it.

// Free Cheat Sheet

The one-page version, free

I made the reference I wish I had for that first session. Everything above, plus tiling and a full glossary, on one page you can keep next to the scope. No email required.

Download the cheat sheet → Read the machine vision terms

The machine vision terms I wish someone had just explained to me

I did not come from a research background. Army, then deputy sheriff, then software. So when I started working in imaging, every conversation was full of words I was supposed to already know. Bit depth. Signal to noise. Spatial resolution. People said them like they were obvious, and I spent a lot of nights quietly googling terms and still feeling lost. This is the reference I wish I had back then.

01A pixel is the smallest unit

This was the first thing that tripped me up. Every digital image is just a grid of tiny squares, and each square is one color. Resolution is how many squares you have. More pixels means more detail, but also bigger files and slower processing.

The part that confused me longest was that resolution means two different things in microscopy. Pixel resolution is how many pixels are in the image. Optical resolution is the smallest distance the microscope can actually separate. They are not the same, and mixing them up cost me a few embarrassing questions. A good rule is to sample each smallest feature with at least two pixels.

02Bit depth is how much each pixel can hold

Bit depth controls how many brightness levels a single pixel can record. An 8 bit image holds 256 levels. A 16 bit image holds 65,536. It took me a while to understand why anyone cared.

Here is the version that finally clicked for me. Low bit depth throws away information you can never get back. In fluorescence that can be the difference between catching a faint structure and losing it in the rounding. For anything you plan to publish or measure, 16 bit is almost always the right call.

03Get exposure right before anything else

Exposure is just how long the sensor sits open collecting light. Too long and the bright areas clip to pure white, and that detail is gone for good. Too short and your real signal disappears into the noise. With live cells, long exposures also stress the sample with phototoxicity and bleaching.

The habit that saved me was checking the histogram before saving every time. A big spike jammed against the right edge means you already lost data, and no amount of editing brings it back.

04Signal, noise, and picking a sensor

Signal is the thing you actually want to see. Noise is all the random junk that comes from the sensor and from light itself. A high signal to noise ratio is easy to read. A low one is buried and frustrating. Something nobody told me early on is that you can often clean it up by averaging a few frames instead of just blasting more light at it.

On sensors, sCMOS is the default for most fluorescence work now. EMCCD only earns its higher cost when you genuinely have just a handful of photons per pixel.