Fine-Tuning Autofocus in NINA
Most people set up autofocus in NINA once, watch it find a focus point, and never think about it again. That's fine right up until the night it isn't. A run takes forever, or a curve comes out lopsided, or that yellow warning appears telling you the new focus is worse than the old one. Then you start poking at settings, and it's easy to make things worse before you make them better.
This is a guide to tuning autofocus deliberately rather than by trial and error. The goal isn't a textbook-perfect curve. It's a curve you can trust, produced in a reasonable amount of time, night after night.
Start by reading the curve
Before you touch a single setting, learn to read the result you're already getting. Every autofocus run in NINA gives you a plot and a set of numbers, and those tell you almost everything you need to know.
The number to anchor on is R², the goodness-of-fit. This is how well the curve NINA drew matches the points it measured. Anything at 0.9 or above is a solid fit. Down around 0.85 and lower, the run is getting noisy and the reported focus position is less reliable. If you're consistently seeing high R² values, your focus is being found correctly, full stop. A lot of “problems” people chase are cosmetic: the curve looks a bit ugly, but the fit is fine and the focus is genuinely good.
The second thing to look at is the shape. You want a clear V or U with a well-defined bottom, and ideally several measurement points clustered near the minimum rather than one lonely point at the bottom and everything else piled up the sides. A densely sampled minimum gives the curve-fit something to work with and produces a focus position you can rely on.
The error bars matter too. Those vertical red lines show the measurement uncertainty at each point. Short, tight bars mean confident measurements. Long bars mean the run was fighting noise: poor seeing, thin cloud, low signal, or too few stars. If your bars are consistently large, that's your real problem, and no amount of curve-fitting trickery fixes it. You fix it at the source, which we'll come to.
The settings that actually move the needle
NINA's autofocus page has a lot of fields, but only a handful genuinely change your results. Here's what each one does and when to reach for it.
Step size controls how far the focuser moves between measurements. Too small and every point sits near the bottom of the curve, so the fit has no leverage and the whole thing is mushy. Too large and you overshoot the useful range, with your outer points shooting up into huge, uninformative HFR values. You want a step size where the outermost points land at roughly two to three times the in-focus HFR. That gives a nicely proportioned curve without wasting time out in the weeds. If your outer points are up at four or five times the minimum, your step size (or your offset count) is too aggressive.
Initial offset steps sets how many points get measured on each side of the estimated focus. More points make for a more robust fit but cost time. Fewer points are faster and, if your fits are already landing at high R², perfectly adequate. Three or four per side is the usual sweet spot. Dropping from four to three is one of the easiest ways to shave time off a run once you've confirmed your fits are healthy.
Exposures per point is where runs get expensive. Set this to two and every single position now costs two frames, two downloads, and two star-detection passes. Averaging multiple exposures does smooth out seeing and tighten those error bars, so it's a real quality lever, but it's also the most costly one you have. Treat it as something to spend deliberately, not a default.
Exposure time determines how much signal each measurement gets. Longer exposures mean brighter, more detectable stars and better measurements, especially through narrowband filters where stars are dim and sparse. The trade is obvious: longer exposures make longer runs.
There's a useful interaction between those last two. One longer exposure often beats two shorter ones. A single frame at, say, six or seven seconds gathers similar signal to two frames at four, but with half the downloads and half the detection overhead. If you're running two exposures per point mostly to claw back signal-to-noise, try collapsing that into one longer exposure instead and compare the R² and the run-to-run consistency. You'll frequently get the same fit quality for less time.
When one side of the curve looks steeper
This one throws people constantly. You get a curve where one arm climbs steeply and the other rises gently, and it looks like something's broken. Usually nothing is.
A truly symmetric focus curve only exists in an ideal optical system. In the real world, residual spherical aberration, a fast focal ratio or slightly imperfect collimation all make the star blur grow at different rates on either side of focus. That asymmetry is baked into your optics, not your focuser. It shows up as uneven arm slopes and it's completely normal.
Noise makes it worse. Low signal-to-noise per point, the kind you get in narrowband or poor seeing, exaggerates whatever asymmetry is already there and varies it run to run. Heavily defocused stars behave strangely too: faint ones drop below detection, bright ones bloom into donuts if you've got a central obstruction, and the far ends of the curve stop behaving in a tidy, mirror-image way.
The takeaway: mild-to-moderate asymmetry is expected and not something you fix by fiddling with mechanical settings. Very strong, persistent asymmetry can be a gentle hint to check your collimation, but it's a hint, not a diagnosis. Don't go chasing backlash or motor settings to make a curve symmetrical, because that's not what's causing it.
About backlash
Backlash deserves its own section because it's the most misunderstood setting in the whole autofocus workflow, and it's the wrong tool for most of the problems people point it at.
The key insight
When NINA measures the curve, it approaches every single point from the same direction. It racks out to the start and then steps steadily inward through all the positions. Because every measurement is approached the same way, backlash cannot distort the shape of the curve. It can't make one arm steeper than the other. So if your curve looks asymmetric, backlash is not the culprit, and adding backlash compensation won't straighten it out.
What backlash does affect is the final move. After the curve is measured and the optimum position is computed, the focuser has to travel back to that mid-curve point, and that reversal runs in the opposite direction to all the measurement steps. If your focuser has real mechanical backlash, that reversal can land you slightly short of the true optimum, leaving you a touch out of focus. This is where compensation earns its keep: not in the curve, but in landing accurately on the computed position and in run-to-run repeatability.
Many modern electronic focusers have very little backlash to begin with. If you measure yours and it comes out near zero, running with zero is entirely legitimate and you're not missing anything. If you do have measurable backlash, put the value in and use the Overshoot compensation method, which tolerates a generous overestimate far better than the Absolute method does. When Overshoot is selected but the value is left at zero, it simply does nothing, which is exactly the right behaviour for a low-backlash focuser.
To measure it: rack the focuser well out, then step back in and note where the HFR actually starts changing. The lag between when you commanded movement and when focus started responding is your backlash. It's often smaller than people expect.
Let the plugin do the hard part
If you're not already running the Hocus Focus plugin, it's worth installing purely for two settings that address the exact problems above.
Uneven Hyperbolic Fit is the direct answer to an asymmetric curve. A standard hyperbola is symmetric by definition, so it forces both arms to mirror each other and gets dragged off-centre when your real data is lopsided. The uneven variant lets each arm take its own curvature, so it models the actual shape instead of splitting the difference. The plotted curve will still look asymmetric, because your data genuinely is, but the computed minimum now reflects reality. This is a cleaner solution than switching NINA to a trend-based fitting strategy.
Weighted Hyperbolic Fit is the companion for noisy nights. It weights each measurement by its uncertainty, so the tight, confident points pull hard on the fit and the noisy ones with big error bars pull less. On a low-signal narrowband run this does real work, stopping one bad, uncertain point from yanking your focus position around.
Keep the fitting strategy consistent
These hyperbolic options only engage if NINA's own curve-fitting strategy is set to Hyperbolic. If you switch NINA to Parabolic or a trend fit, the uneven and weighted logic won't apply. Set them together and leave them.
The rest of Hocus Focus is sensible out of the box. Validating HFR improvement with a reasonable tolerance is a good safety net against a bad run, and allowing a small number of outlier rejections lets the fit throw out the occasional point ruined by a satellite trail or a gust of bad seeing.
Quality versus time
Every quality lever costs time, and it's easy to over-tune to the point where autofocus eats a chunk of your imaging night. Keep some perspective here.
A run at R² 0.98 and a run at R² 1.00 will land you at focus positions that differ by an amount well inside the noise. The perfect-looking curve isn't meaningfully better focus. It's just smoother and a little more reliable. That means you have room to trim. If a run got slow because you pushed exposures per point up and exposure time out, you can usually pull one of those back and lose almost nothing.
The bigger lever, though, isn't per-run cost at all. It's how often you refocus. A three-minute autofocus once every degree or two of temperature change across a night is trivial overhead. The same run fired every twenty minutes on a timer is where autofocus actually starts hurting your session. Trigger refocus on a temperature delta or on measured HFR drift rather than a tight time interval, and the per-run cost mostly stops mattering.
That “worse than the previous” warning
Sooner or later NINA will pop a warning that the new focus point's HFR is significantly worse than the previous one, and reject the run. Don't panic. Read it properly, because it's telling you something specific.
The clue is in the comparison. If the curve fit itself was good (high R²), then the fit found the right position. The problem opened up after the curve was computed, on the validation frame taken at the final position. That narrows the cause to two suspects.
The first is transient conditions. A validation frame is usually a single exposure, and one unlucky moment of bad seeing or a wisp of cloud reads high. This is random and won't repeat. The second is backlash on that final approach: the reversal we talked about earlier landing you slightly short of the computed optimum, so the validation frame genuinely is a little soft.
You tell them apart by frequency. Run autofocus a few more times. If the warning fires most runs with a similar predicted-versus-measured gap, it's systematic, so look at backlash. If it's a one-off and subsequent runs validate cleanly, it was conditions, and you can forget it. Either way, if the curve fit was good, the run found the right place, and only the validation frame was in question.
How often to refocus
Set your refocus triggers and then leave autofocus alone for most of the night. The two triggers worth using are temperature change and HFR drift. As your gear cools through the evening the focus point moves, and a temperature trigger tracks that naturally. An HFR-drift trigger catches anything the temperature trigger misses, whether that's flexure, a filter change, or something else. Between them you refocus when you actually need to and not on an arbitrary clock. That's the setup that keeps sharp stars all night without autofocus dominating your imaging time.
The short version
Tune from the data, not from superstition. Read R² and the curve shape before you change anything. Spend your time budget on signal-to-noise, meaning exposure time and, when needed, exposures per point, because clean measurements fix more than any clever setting. Let the uneven and weighted hyperbolic fits handle curve asymmetry and noise. Treat backlash as a final-approach and repeatability tool rather than a curve-shape fix, and don't be afraid to run it at zero if your focuser is well-behaved. And refocus on temperature and drift, not on a timer.
Get those right and autofocus becomes something you set once and stop thinking about, which is exactly where you want it.