Super-Resolution Microscopy: SIM vs STED vs SMLM

The decision between SIM, STED and single-molecule localization microscopy (SMLM) is almost never settled by the resolution number in the brochure. It is settled by two things the brochure does not mention: whether your label can physically support the resolution you are asking for, and whether the resulting image can be shown not to be a reconstruction artefact. Groups that book time on a 20 nm instrument and hand it a sparsely labelled, indirectly immunostained sample get a 20 nm instrument producing a picture whose real information content is nowhere near 20 nm — and, because all three methods generate their images computationally, one that can look convincing anyway.

Start from the structure you need to resolve, the label you can realistically attach to it, and whether the sample has to be alive. The instrument is the last choice, not the first.

What each method does, and what it actually measured

All three bypass Ernst Abbe’s diffraction limit of roughly 0.2 µm, the achievement recognised by the 2014 Nobel Prize in Chemistry awarded jointly to Eric Betzig, Stefan W. Hell and William E. Moerner “for the development of super-resolved fluorescence microscopy”. They bypass it by completely different routes, and those routes are what determine which one suits your experiment.

SIM (structured illumination microscopy) illuminates with a patterned grid at several angles and phases, then computationally unmixes the resulting moiré fringes to recover spatial frequencies the objective cannot pass. Linear SIM buys roughly a factor of two in every direction — approximately 100–120 nm laterally and 350–400 nm axially.
STED (stimulated emission depletion) is a scanning confocal method with a second, doughnut-shaped depletion beam that switches off fluorescence everywhere except a sub-diffraction spot at the doughnut’s centre. Reported ranges are roughly 30–50 nm laterally and 100–200 nm axially, and the resolution is tunable in real time by raising depletion power.
SMLM (PALM, STORM, dSTORM and relatives) makes almost all fluorophores dark at any instant, localises the few that blink on to a precision far below the width of their point spread function, and builds the image from tens of thousands of frames. Localization precision is typically ~20–50 nm or better, with lateral resolution figures commonly quoted around 20–50 nm and axial around 10–70 nm depending on the 3D scheme.

Those ranges come from different papers, different samples and different definitions of “resolution”, which is exactly why they are hard to compare. The most useful single data point in the literature is a study that put the same structures, prepared in parallel, through all three. Wegel and colleagues measured the full width at half maximum of labelled microtubules — a filament about 25 nm across, so essentially a resolution test object — and of the centriole wall:

Microtubules: SIM 107 (±0.85) nm; STED 58.8 (±1.2) nm; SMLM 56.3 (±0.78) nm.
Centriole wall: SIM 123 (±1.2) nm; STED 83.5 (±1.9) nm; SMLM 81.4 (±1.7) nm.

Two things follow. First, STED and SMLM landed within a few nanometres of each other on real biological specimens, despite the very different theoretical ceilings people quote for them — on ordinary immunostained samples the label, not the physics, was the binding constraint. Second, SIM was roughly a factor of two behind, which is exactly what linear SIM promises and no more. If you need to distinguish two objects 60 nm apart, SIM will not do it and no amount of reconstruction tuning will change that.

The constraint most people ignore: the label

Resolution in fluorescence microscopy is not a property of the microscope alone. Three label-side effects routinely dominate, and all three are decided before the sample reaches the stage.

Labelling density

You cannot resolve a feature you have not sampled. The relevant rule is the Nyquist criterion applied to fluorophore spacing, and it is more demanding than most people assume: the SMLM Primer notes that fivefold higher sampling than plain Nyquist is actually required to support a given resolution, and that with a localization precision of 10 nm, 22% of molecules must be localized to achieve a resolution of 23 nm. Under-labelled structures produce images that are punctate rather than continuous, and it is very easy to over-interpret that punctate appearance as biological clustering.

This is also why some structures are much harder than others regardless of technique. Wegel and colleagues point out there are roughly six times as many binding sites per unit length on microtubules as on actin, which is a large part of why actin networks are so much less forgiving.

Linkage error

An antibody does not sit on the epitope; it holds the dye some distance away. The SMLM Primer puts this at ~10 nm, or up to ~20 nm for indirect immunolabelling with a primary plus a labelled secondary. If your target resolution is 20–30 nm, a two-antibody sandwich has already spent most of your error budget before the laser is switched on. Nanobodies, direct primary conjugates or genetically encoded tags are the fix, and they need to be planned into the experiment, not retrofitted.

Photon budget

For SMLM the theoretical floor is set by photon count — localization uncertainty scales as σ₀/√N, and with σ₀ around 100 nm and N between 10² and 10⁴ photons the floor is 1–10 nm. That the practical figure is 20–50 nm is the gap between physics and labelling reality. It also explains why dye choice is not cosmetic: Alexa Fluor 647 dominates dSTORM work specifically because of its high emitted photon count.

Sample preparation is not interchangeable between the three

This is the most common practical failure when a lab moves a project from one modality to another. A slide prepared for SIM is often unusable for SMLM.

Fixation. Crosslinking fixation (commonly paraformaldehyde) is required across the board; methanol and methanol-containing formalin should be avoided, because they degrade the fine structure you are trying to resolve.
Fluorophores. SIM is fully compatible with essentially any commercially available fluorophore — a genuine practical advantage. STED needs bleaching-resistant dyes matched to the available depletion line. SMLM needs photoswitchable dyes, which is a much narrower list.
Mounting medium. SIM and STED work are typically mounted in a hard-setting antifade such as Prolong Gold. dSTORM requires an imaging buffer — PBS with 10–100 mM of a thiol such as mercaptoethylamine plus an enzymatic oxygen scavenger (glucose oxidase, catalase and glucose) — which is prepared fresh, is short-lived, and cannot coexist with a cured mountant. Refractive-index matching matters throughout, and mismatch is a leading cause of SIM artefacts.
Optics. All three depend on a well-corrected high-NA objective and a coverslip of the correct thickness; see microscope objective selection. Pixel size must be calibrated per objective before any distance measurement is claimed — see calibrating with a stage micrometer.

Acquisition time, laser dose and live cells

The three methods are separated by more than an order of magnitude in how long they occupy the stage and how hard they hit the sample. In the parallel comparison, SIM z-stacks took up to a few minutes, STED took seconds to minutes, and SMLM ran up to twenty minutes to accumulate tens of thousands of frames. The SMLM Primer gives 10,000–20,000 frames at 10–100 ms exposure, or 2–30 minutes, as typical for a dense structure.

Dose is the mirror image. STED is the aggressive one: depletion intensities on the order of 150 MW/cm² in the focal plane were reported in that study, with up to 260–300 mW at the objective back aperture. SIM and SMLM operate in a linear, lower-intensity regime, though SMLM compensates with a very long total exposure.

For live imaging this ordering matters more than resolution does:

SIM is the default for live super-resolution. It works with ordinary fluorescent proteins, needs no special buffer and no punishing laser power. Algorithmic advances have pushed it a long way: Hessian-SIM reconstructs artefact-minimised images using less than 10% of the photon dose of conventional SIM and reached a spatiotemporal resolution of 88 nm at 188 Hz, enabling hour-long time-lapse super-resolution imaging of actin in live cells.
STED is viable live but on a photobleaching clock, and its usable duration is set by how much depletion power your specimen and fluorophore tolerate.
Live SMLM is possible but constrained: the acquisition time needed for a dense reconstruction is long relative to most cellular dynamics, so you are usually choosing between a well-sampled image and a moving one.

Thickness, depth and the shape of your specimen

SIM is a widefield method and works best on thin, low-scattering samples, because out-of-focus light degrades the moiré signal it depends on. Counterintuitively, though, its optical sectioning makes it the better choice for dense meshworks and structures whose signal extends along the z-axis, where SMLM struggles to keep single emitters isolated. STED, being confocal, tolerates thicker samples better. SMLM is at its best on isolated, near-coverslip structures — which is why it pairs so naturally with TIRF illumination, whose thin evanescent excitation field suppresses background from deeper in the cell.

Depth degrades all three, and it degrades STED disproportionately, because the resolution depends on the quality of the zero-intensity minimum at the centre of the depletion doughnut and aberration fills that minimum in. Note too that a standard 2D STED doughnut sharpens only laterally: axial performance stays at the confocal level unless the system implements a 3D depletion pattern, which improves z at the cost of lateral resolution and of much greater sensitivity to aberration. If your question is genuinely three-dimensional, ask the facility which depletion geometry the instrument has before you assume the quoted resolution applies in z.

A decision sequence that works

State the separation you need to resolve, in nanometres. If it is above ~120 nm, SIM is sufficient and is the cheapest, fastest, least destructive option. Do not buy resolution you do not need.
Check that your label can support it. Below about 50 nm, an indirect immunostain with its ~20 nm linkage error is a design flaw. Fix the label before choosing the microscope.
Decide whether the sample must be alive. If yes, SIM first, STED second, SMLM only for slow processes or single-particle tracking.
Look at the structure’s density and geometry. Dense 3D meshwork favours SIM or STED; sparse, near-coverslip, countable objects favour SMLM.
Check multicolour requirements early. SIM is the most permissive. STED channels must share a depletion line or the system must carry a second one. SMLM multicolour requires two dyes that switch in the same buffer, which is a real constraint.
Ask what the facility can actually validate. The authors of the parallel comparison put it plainly: it is advisable to use an imaging facility or specialist lab to get the most from super-resolution experiments. This is not a technique to self-teach against a deadline.

Proving the image is real

Every one of these images is computed, and every reconstruction can manufacture structure that was never in the sample. Reviewers increasingly expect evidence, not assertion. Build these into the protocol rather than treating them as an afterthought.

Measure resolution, do not quote the brochure

Report resolution measured from your own data. Fourier ring correlation (FRC) is the standard approach for SMLM. A biological ground truth — the nuclear pore complex is the canonical one — provides an independent check that you can genuinely separate features at the distance you claim.

Run the SIM-specific checks

SIMcheck, an ImageJ toolbox, formalised eight checks across raw and reconstructed 3D-SIM data. On the raw side: a channel intensity profile that exposes bleaching and angle-to-angle variation; a Fourier projection that confirms clean first- and second-order spots, meaning the illumination pattern was actually generated; a motion and illumination variation check; and modulation contrast-to-noise (MCNR), which also sets the appropriate reconstruction filter constant (the paper suggests a Wiener constant w = 0.17/MCNR²). On the reconstructed side: an intensity histogram whose min-to-max ratio separates real features from reconstructed noise; a modulation contrast map that flags apparent “features” sitting on very low underlying MCNR; Fourier plots of the frequency support; and a spherical aberration mismatch check that catches refractive-index problems. The classic SIM artefacts — honeycomb or hammerstroke patterning from a badly chosen Wiener parameter, echoes from spherical aberration, local damage from saturated pixels — are all detectable this way, and all invisible to the naked eye in the final image.

Compare against the diffraction-limited reference

The most general validation applies to any modality. SQUIRREL (Culley et al., Nature Methods, 2018) convolves the super-resolution image back down to diffraction-limited scale and compares it, pixel by pixel, with the widefield reference acquired from the same field. It returns a resolution-scaled error (RSE) and a resolution-scaled Pearson coefficient (RSP), plus a spatial map of where the reconstruction disagrees with the raw data. If a structure in the super-resolution image has no counterpart in the diffraction-limited image of the same field, it was invented. This one check catches a large fraction of publishable-looking nonsense.

The option people forget: expand the sample instead

If none of the three is available — or the facility queue is months long — physically enlarging the specimen is a serious alternative. Standard protein-retention expansion microscopy swells the sample about fourfold in each dimension, giving roughly fourfold better than the diffraction limit, or ~70 nm, on an ordinary microscope. The X10 protocol reaches ~10× linear expansion (over 1,000-fold in volume) and 25–30 nm resolution on conventional epifluorescence microscopes, which its authors argue is comparable to or better than STED and STORM for multicolour work. The catch is that resolution now depends on the expansion being isotropic, which has to be demonstrated rather than assumed — see the guide to choosing an expansion protocol and proving isotropy. Note also that expansion and optical super-resolution are combinable, not mutually exclusive.

Frequently asked questions

Which super-resolution technique has the best resolution?

On paper, SMLM, with reported lateral figures around 20–50 nm and STED close behind at 30–50 nm. In practice the gap narrows or vanishes: measured on identically prepared microtubules, STED gave 58.8 nm and SMLM 56.3 nm. Newer methods such as MINFLUX push further still by combining a patterned excitation minimum with localization, but they are specialist instruments rather than general facility equipment. For most immunostained samples the label, not the instrument, sets the limit.

Is SIM really super-resolution?

Yes, but only by a factor of about two — roughly 100–120 nm laterally against a ~200 nm diffraction limit. That is a genuine and often sufficient improvement, and SIM buys it with standard dyes, low light dose and short acquisitions. It is the wrong tool for anything requiring separation below about 100 nm.

Can I use my existing immunofluorescence slides?

Possibly for SIM, sometimes for STED if the dyes match the depletion line, and almost certainly not for SMLM, which needs photoswitchable dyes and a fresh thiol-plus-oxygen-scavenger imaging buffer that is incompatible with cured hard-set mountant. Also check the fixative: methanol-based fixation should be avoided for all three.

Why does my super-resolution image look punctate or beaded?

Most often under-labelling. Supporting a given resolution requires roughly fivefold denser fluorophore sampling than plain Nyquist, and a sparsely labelled continuous filament reconstructs as a string of dots. Increase labelling density, or reduce the resolution you claim, before interpreting the beading as clustering.

Which technique works best for live cells?

SIM, in most cases. It tolerates ordinary fluorescent proteins, needs no exotic buffer, and modern reconstruction algorithms have cut its dose dramatically — Hessian-SIM achieved 88 nm at 188 Hz using under 10% of conventional SIM’s photon dose, supporting hour-long time lapses. STED is live-capable but limited by depletion power and bleaching; SMLM’s long acquisitions suit slow processes and single-particle tracking rather than general live imaging.

How do I convince a reviewer the structure is not an artefact?

Report a measured resolution (FRC for SMLM), show a diffraction-limited widefield image of the same field alongside the reconstruction, and run a quantitative comparison such as SQUIRREL’s RSE and RSP. For SIM, include the SIMcheck raw-data diagnostics — modulation contrast in particular — so the reviewer can see the reconstruction was supported by the raw frames.

How much does labelling strategy matter compared with instrument choice?

Below roughly 50 nm it matters more. Indirect immunolabelling introduces up to ~20 nm of linkage error before any imaging occurs; a direct primary conjugate, a nanobody or a genetically encoded tag removes most of it. Choosing a better microscope while keeping a two-antibody sandwich rarely improves the result.