GLIDE: Symmetry, Diffusion, and Robotics

• GLIDE is a multifaceted concept that, in physics, denotes a nonsymmorphic symmetry combining reflection with fractional translation to control band connectivity and topological invariants.
• In machine learning, GLIDE refers to a guided language-to-image diffusion system that uses a two-stage process for photorealistic image synthesis and inpainting.
• Additional applications of GLIDE include quadrupedal locomotion and speculative decoding in LLMs, highlighting its versatility across diverse research fields.

GLIDE denotes several distinct technical constructs across contemporary research. In condensed-matter physics, photonics, acoustics, and wave mechanics, “glide” usually refers to a nonsymmorphic symmetry: a reflection combined with a fractional translation, or, in spatiotemporal settings, a reflection combined with a half-period time translation. In that role it controls band connectivity, topological invariants, interface states, and symmetry-enforced selection rules (Lei et al., 2023). In machine learning, GLIDE is also the acronym for “Guided Language to Image Diffusion for Generation and Editing,” a text-conditional diffusion system for photorealistic image synthesis and inpainting (Nichol et al., 2021). Related acronymic variants include GLiDE for quadrupedal locomotion and GliDe for speculative decoding acceleration (Xie et al., 2021, Du et al., 2024).

1. Glide as a nonsymmorphic symmetry operation

A glide operation is a reflection followed by a fractional lattice translation. In a glide-symmetric phoxonic crystal waveguide, the operator is written

$\hat{G}\psi(x,y)=\psi(x+a/2,-y),$

so the symmetry is neither pure reflection nor pure translation, but their combination (Lei et al., 2023). In magnetic photonic crystals, an analogous form is

$G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$

namely reflection in the $y$-direction followed by a half-translation along $z$ (Kim et al., 2021). In spinful Bloch problems, the square of the glide on glide-invariant states is momentum dependent,

$$g^2 = -e^{-ik_yR_2}, \qquad \Delta_{\pm}(k_y)=\pm i\,e^{-ik_yR_2/2},$$

which is the characteristic nonsymmorphic phase structure absent for ordinary mirror symmetry (Alexandradinata et al., 2019).

This momentum dependence is operationally decisive. In spin- and angle-resolved photoemission, the glide eigenvalue of an initial Bloch state contains the factor $e^{-ik_z c/2}$, and that phase changes sign when the photoelectron momentum is shifted by one surface reciprocal lattice vector. The resulting spin-selection rule therefore depends not only on polarization but also on the surface Brillouin zone index, which distinguishes glide symmetry sharply from mirror symmetry (Ryoo et al., 2018).

A spatiotemporal analogue appears in time-modulated waveguides. There, temporal glide combines mirror reflection $P_z:z\mapsto -z$ with a half-period time translation. If the one-period Floquet operator is $M=U_B U_A$ and the half-period maps satisfy $P_z U_A = U_B P_z$, then

$G_t=P_z U_A, \qquad G_t^2=M.$

In this setting, glide squares to Floquet time evolution itself rather than to a spatial translation, so its primary consequence is not ordinary spatial “band sticking” but a constraint on the sideband structure of Floquet eigenstates (Camacho, 11 Jun 2026).

2. Guided waves, dispersion engineering, and slow-light regimes

Glide symmetry has been used to create guided states inside bulk gaps without relying on the standard Dirac-point-opening mechanism. In a square-lattice phononic crystal of steel rods in water, a domain wall is formed by shifting one half-crystal by a distance $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$0 along the interface direction. For $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$1, the two halves are aligned and the structure has a complete band gap with no transmission. For $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$2, the interface becomes glide-reflection symmetric and supports guided modes inside the complete band gap. The associated topological mismatch is expressed by a 2D Zak-phase shift

$G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$3

so that $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$4 at $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$5. The experiment reported transmission from approximately $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$6 MHz to $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$7 MHz, while the complete bulk band gap extended from about $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$8 MHz to $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$9 MHz; for $y$0 and $y$1, a mini-gap opened and the transmission window narrowed (Martínez et al., 2022).

A closely related mechanism was implemented in a 2D silicon-slab phoxonic crystal waveguide split into two identical sub-crystals across a glide plane. In the glide-symmetric case $y$2, the antiunitary combination $y$3 satisfies

$y$4

so at the Brillouin-zone boundary $y$5, $y$6, yielding a Kramers-like degeneracy. The super-cell bands therefore degenerate in pairs at the zone edge, and gapless interface-localized guided modes appear in photonic and phononic band gaps for TM, TE, in-plane elastic, and out-of-plane elastic excitations. The paper identified broad single-mode windows, including $y$7 for TM and $y$8 for TE. When the glide dislocation is tuned from $y$9 to $z$0, the edge band gaps shrink continuously to zero; at $z$1, the guided bands become more weakly dispersive and approach flat single-mode guided bands (Lei et al., 2023).

Suspended silicon photonic-crystal waveguides with one side shifted by half a period along the waveguide axis supply an optical slow-light realization. Resonator measurements and evanescent coupling were used to reconstruct the dispersion, and group indices exceeding $z$2 were measured, with a maximum predicted group index of about $z$3. The data also showed that the glide-enforced zone-edge degeneracy does not coincide with the group-index maximum, and that fabrication-induced backscattering becomes increasingly important in the high-$z$4 regime, particularly for longer devices (Patil et al., 2021).

Glide symmetry also reshapes effective-medium behavior in metallic metamaterials. In a metallic parallel-plate structure with glide-symmetric elliptical holes, the glide configuration corresponds to a half-period displacement $z$5 between the two perforated plates. The resulting parity-dependent coupling suppresses the stopband between the first and second modes, reduces the dispersion of the fundamental mode, and produces an anisotropic refractive index over a wide frequency range. Combined with transformation optics, this enabled a compressed Maxwell fish-eye lens with $z$6 size compression operating from $z$7 GHz to $z$8 GHz; the paper further reported a maximum achievable compression of about $z$9 in the glide case, compared with only $$g^2 = -e^{-ik_yR_2}, \qquad \Delta_{\pm}(k_y)=\pm i\,e^{-ik_yR_2/2},$$0 in the non-glide version (Alex-Amor et al., 2021).

In acoustics, a glide-symmetric corrugated waveguide is obtained when $$g^2 = -e^{-ik_yR_2}, \qquad \Delta_{\pm}(k_y)=\pm i\,e^{-ik_yR_2/2},$$1 and $$g^2 = -e^{-ik_yR_2}, \qquad \Delta_{\pm}(k_y)=\pm i\,e^{-ik_yR_2/2},$$2. In that limit the two guided branches touch and no band gap exists between them; the upper branch can exhibit a negative slope while remaining forward propagating because the effective period is halved. This slow-wave behavior was used for sensing and interferometric isolation. For a water–glycerol sensing platform, the paper reported a transmission-zero sensitivity of about $$g^2 = -e^{-ik_yR_2}, \qquad \Delta_{\pm}(k_y)=\pm i\,e^{-ik_yR_2/2},$$3 Hz/\% glycerol and a phase-shift sensitivity of about $$g^2 = -e^{-ik_yR_2}, \qquad \Delta_{\pm}(k_y)=\pm i\,e^{-ik_yR_2/2},$$4 rad/\% glycerol. A Mach–Zehnder acoustic isolator built from two such glide-symmetric arms achieved isolation greater than $$g^2 = -e^{-ik_yR_2}, \qquad \Delta_{\pm}(k_y)=\pm i\,e^{-ik_yR_2/2},$$5 dB (Janković et al., 2020).

3. Topological classification and symmetry indicators

In gapped photonic systems, glide symmetry supports $$g^2 = -e^{-ik_yR_2}, \qquad \Delta_{\pm}(k_y)=\pm i\,e^{-ik_yR_2/2},$$6-classified topological phases. For magnetic photonic crystals in space group No. 230, $$g^2 = -e^{-ik_yR_2}, \qquad \Delta_{\pm}(k_y)=\pm i\,e^{-ik_yR_2/2},$$7, the glide-$$g^2 = -e^{-ik_yR_2}, \qquad \Delta_{\pm}(k_y)=\pm i\,e^{-ik_yR_2/2},$$8 invariant is given by

$$g^2 = -e^{-ik_yR_2}, \qquad \Delta_{\pm}(k_y)=\pm i\,e^{-ik_yR_2/2},$$9

and simplifies to

$e^{-ik_z c/2}$0

where $e^{-ik_z c/2}$1 is the number of photonic bands below the gap. Consequently, the gap between the second and third lowest bands is always topologically nontrivial whenever it is opened, because $e^{-ik_z c/2}$2 implies $e^{-ik_z c/2}$3. In the time-reversal-symmetric case, the second and third bands meet at the $e^{-ik_z c/2}$4 point in a fourfold generalized Dirac point; staggered magnetization breaks time-reversal symmetry while preserving glide and opens the topological gap, whereas uniform magnetization does not, due to minimal band connectivity exceeding two (Kim et al., 2021).

For three-dimensional time-reversal-invariant solids, glide symmetry refines the usual strong-topological-insulator classification. The resulting strong invariant becomes $e^{-ik_z c/2}$5, and for glide-symmetric Weyl metals the classification extends to $e^{-ik_z c/2}$6. The $e^{-ik_z c/2}$7 factor counts the net number of Weyl points in a symmetry-reduced quadrant of the Brillouin zone, while the $e^{-ik_z c/2}$8 invariant $e^{-ik_z c/2}$9 is encoded in the Wilson-loop or Zak-band connectivity of glide sectors. Material-specific values reported in the paper include $P_z:z\mapsto -z$0 for Ba$P_z:z\mapsto -z$1Pb, $P_z:z\mapsto -z$2 for stressed Na$P_z:z\mapsto -z$3Bi, and $P_z:z\mapsto -z$4 for KHgSb (Alexandradinata et al., 2019).

Interacting glide-protected topological phases admit an especially concise formulation through a coupled-layer construction. If two-dimensional short-range-entangled phases with on-site symmetry $P_z:z\mapsto -z$5 form an Abelian group $P_z:z\mapsto -z$6, then the corresponding three-dimensional glide symmetry protected topological phases are classified by

$P_z:z\mapsto -z$7

Equivalently, the paper established the universal stacking rule

$P_z:z\mapsto -z$8

This implies that every nontrivial three-dimensional glide-protected phase is effectively $P_z:z\mapsto -z$9-torsion. The same framework yields anomalous surface topological orders that symmetrically gap the surface while remaining impossible in a strictly two-dimensional system with the same symmetry (Lu et al., 2017).

When inversion symmetry is added, the algebra of glide indicators depends on the resulting space group. In space group No. 13, the redefined glide invariant $M=U_B U_A$0 can be written purely in terms of high-symmetry irreducible representations and satisfies $M=U_B U_A$1, so the glide-$M=U_B U_A$2 invariant becomes a genuine symmetry indicator. In space group No. 14, by contrast, high-symmetry irreps determine only the combination $M=U_B U_A$3, not $M=U_B U_A$4 and the Chern number separately (Kim et al., 2018).

A related nonprimitive-lattice generalization was derived for base-centered monoclinic systems. For space group No. 9, the glide-$M=U_B U_A$5 invariant must be reformulated because $M=U_B U_A$6 is not glide invariant in the base-centered Brillouin zone. After Brillouin-zone folding, the invariant is expressed through modified Berry-curvature and Berry-phase integrals. With added inversion symmetry, giving space group No. 15, the glide invariant reduces to a Fu–Kane-like formula and, in the spinless case, to

$M=U_B U_A$7

(Kim et al., 2020).

4. Selection rules, spectroscopy, and synthetic dynamics

Glide symmetry generates experimentally accessible selection rules because its eigenvalues carry momentum-dependent fractional-translation phases. In glide-symmetric spin- and angle-resolved photoemission, with the surface, incident light, and photoelectron momentum all invariant under the glide, the emitted photoelectron spin is fixed by the initial-state glide label $M=U_B U_A$8 and the surface-zone index $M=U_B U_A$9. For $P_z U_A = U_B P_z$0-polarized light,

$P_z U_A = U_B P_z$1

while for $P_z U_A = U_B P_z$2-polarized light,

$P_z U_A = U_B P_z$3

The key consequence is that the spin polarization flips when the photoelectron is shifted by one surface reciprocal lattice vector from the first to the second surface Brillouin zone. Simulations on KHgSb (010) showed precisely this zone-dependent reversal for hourglass-like surface states (Ryoo et al., 2018).

The same momentum dependence underlies a more general photoemission program for measuring glide-resolved topological invariants. In glide-symmetric solids, the plane-wave photoelectron satisfies

$P_z U_A = U_B P_z$4

which produces an alternating fully spin-polarized “fan” of emitted rays as the photoelectron momentum changes by reciprocal-lattice increments. This spectroscopic spin-momentum locking is the proposed route to reading out the $P_z U_A = U_B P_z$5 glide invariant of a surface Bloch state and, in ideal conditions, to generating a source of fully spin-polarized photoelectrons (Alexandradinata et al., 2019).

Temporal glide produces a different selection rule. Writing a Floquet eigenstate as

$P_z U_A = U_B P_z$6

the temporal-glide condition implies

$P_z U_A = U_B P_z$7

Thus even sidebands preserve transverse parity and odd sidebands flip it. In the scalar time-modulated trilayer waveguide studied in the paper, synchronous modulation gave parity-pure states with $P_z U_A = U_B P_z$8, whereas time-glide modulation produced strongly mixed Floquet states with median $P_z U_A = U_B P_z$9. The wrong-parity violation metric at the glide point reached only $G_t=P_z U_A, \qquad G_t^2=M.$0, and in a finite-section scattering example an incident odd $G_t=P_z U_A, \qquad G_t^2=M.$1 mode converted to the even $G_t=P_z U_A, \qquad G_t^2=M.$2 mode at sideband $G_t=P_z U_A, \qquad G_t^2=M.$3 with $G_t=P_z U_A, \qquad G_t^2=M.$4 (Camacho, 11 Jun 2026).

Glide symmetry also enters synthetic-dimension dynamics. In a spin-dependent double-well optical lattice, glide is realized as half-lattice translation combined with spin exchange, and in the complex-coupling case it becomes a synthetic glide symmetry acting jointly in space and the pumping parameter $G_t=P_z U_A, \qquad G_t^2=M.$5. The symmetry protects band-touching points at $G_t=P_z U_A, \qquad G_t^2=M.$6 and forces the topology of the two lowest bands to be described by a non-Abelian Berry curvature and Wilson line rather than by an Abelian Berry phase. For the inseparable synthetic-glide case, the Wilson line along the $G_t=P_z U_A, \qquad G_t^2=M.$7-direction takes the explicit form

$G_t=P_z U_A, \qquad G_t^2=M.$8

showing coherent mixing between the two low-energy states (Chen et al., 2019).

5. GLIDE as a diffusion model for image generation and editing

In machine learning, GLIDE stands for “Guided Language to Image Diffusion for Generation and Editing.” It is a text-conditional diffusion system built as a two-stage pipeline: a base diffusion model at $G_t=P_z U_A, \qquad G_t^2=M.$9 resolution followed by an upsampling diffusion model from $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$00 to $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$01. The visual component of the base model has about $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$02B parameters, the text Transformer adds about $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$03B parameters, and the total base model is about $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$04B parameters; the upsampler is about $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$05B parameters (Nichol et al., 2021).

The diffusion backbone follows the standard Gaussian forward process

$G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$06

with a learned reverse process

$G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$07

and the usual noise-prediction objective

$G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$08

The central methodological comparison in the paper is between CLIP guidance and classifier-free guidance. For classifier-free guidance, the guided noise estimate is

$G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$09

The paper reported that classifier-free guidance was preferred by human evaluators over CLIP guidance for both photorealism and caption similarity (Nichol et al., 2021).

Guidance	Photorealism Elo	Caption Elo
Unguided	-88.6	-106.2
CLIP guidance	-73.2	29.3
Classifier-free guidance	82.7	110.9

The same study found that samples from the $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$10B text-conditional diffusion model using classifier-free guidance were preferred over DALL-E $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$11 of the time for photorealism and $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$12 of the time for caption similarity, even though DALL-E used CLIP reranking. GLIDE was also explicitly fine-tuned for inpainting by adding four extra input channels—a second RGB image and a mask channel—so that masked regions could be regenerated while preserving context in the unmasked region (Nichol et al., 2021).

A direct downstream use of this inpainting mode was explored for human action-effect prediction. The task is to take an initial-state image and an action phrase and predict an image of the world after the action while preserving the same scene context. The method masks the region where the effect is expected to occur and uses GLIDE to inpaint that region conditioned on text. Three masking strategies were examined: a fixed mask, segmentation masks from Mask R-CNN detections in EPIC-KITCHENS-100, and hand-object detection masks from a Faster R-CNN detector filtered with a significance threshold of $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$13. The prompt itself could be either the raw action phrase or a GPT-3-generated action-effect description based on two randomly selected action/effect pairs from the Gao et al. dataset. The paper reported qualitative improvements from richer effect descriptions and more precise segmentation masks, while also noting failure cases for global appearance changes such as “turn on light,” position-changing actions, and cases where the effect region exceeded the mask or object identity needed to be preserved exactly (Li et al., 2022).

6. Other acronymic uses: locomotion and speculative decoding

A separate acronymic use is GLiDE, “Generalizable Quadrupedal Locomotion in Diverse Environments with a Centroidal Model.” This method trains reinforcement learning on a centroidal rigid-body model with four massless legs rather than on a full-body simulator. The policy outputs desired centroidal accelerations $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$14, and a quadratic program converts those commands into feasible ground reaction forces under stance, swing, and friction constraints. The reported system used $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$15 simulations in parallel, achieved roughly $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$16–$G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$17k steps/s with about $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$18 s per environment step on average, and demonstrated flat-terrain gaits, stepping-stone locomotion, two-legged in-place balance, balance-beam locomotion, and direct sim-to-real transfer (Xie et al., 2021).

Another distinct usage is GliDe in large-language-model inference. “GliDe with a CaPE” modifies speculative decoding by replacing the standard draft model with a “Glimpse Draft Model” that cross-attends to the frozen target model’s cached keys and values. In the paper’s notation, the draft hidden states are projected to queries $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$19, which attend to upper-layer target-model caches. The study reported acceptance-rate improvements ranging from $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$20 to $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$21 percentage points relative to an otherwise similar draft model without KV-cache reuse. In walltime, GliDe accelerated Vicuna models by up to $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$22, and the confidence-aware proposal expansion method CaPE extended the improvement to $G_y=\{M_y\,|\,\tfrac{a}{2}\hat{z}\},$23 (Du et al., 2024).

These acronymic usages are unrelated to the nonsymmorphic symmetry operation, but they retain the same lexical form. The term GLIDE therefore has a strongly polysemous role in current research literature: as a symmetry principle in wave and topological physics, as a diffusion architecture for text-guided image synthesis and editing, and as a compact label for distinct systems in robotics and LLM acceleration.