Source-linked AI summary
Strike a Chord! Modal Kinetic Typography
Maham Tanveer, Jiyeon Han, Nanxuan Zhao, Hao Zhang
TL;DR
Kinetic typography automation must make text expressive while keeping glyphs legible and making motion itself carry meaning. Modal kinetic typography builds motion from glyph-derived modes and uses a frozen video prior to tune their amplitudes and phases. Against Dynamic Typography, it reports less stretch and higher concept similarity, temporal consistency, and motion smoothness.
Problem
Automated kinetic typography must balance expressive, meaningful motion with legibility, including settings where the glyph cannot be reshaped.
Source: §1 · p. 1Method
Results
Against Dynamic Typography, stretch is less than half theirs and lower in 95% of cases, while concept similarity, temporal consistency, and motion smoothness are higher.
Source: §4 · p. 7Takeaways & Limitations
The representation separates shape from motion, allowing a glyph to move without being reshaped.
Source: §6 · p. 9A single outline cannot change topology or turn in depth, and periodic motion excludes actions that end somewhere other than their start.
Source: §6 · p. 9
Abstract
from arXiv · showhide
We introduce modal kinetic typography, which animates a vector glyph to express a semantic concept while keeping it legible. Our key idea is to build motion from the glyph's natural vibration modes. Specifically, a finite-element eigenproblem assembled from the vector outline yields the glyph's softest modes, for the whole letter and for each of its parts, allowing it to bend. The problem's zero-energy solutions, i.e., rigid translations and rotations, are applied in closed form to each part, allowing parts to also move as blocks. To animate the glyph, a frozen video diffusion model supervises only the modes' amplitudes and phases. Our modal approach addresses two weaknesses of prior work. Free-form point optimization under video score distillation (SDS) moves each point and frame independently along noisy gradients, tearing the outline and causing jitter. In contrast, our modes are smooth along the outline and driven by a few whole-cycle harmonics, which restricts these gradients to smooth, seamlessly looping motion. On the other hand, structured alternatives rely on skeletons or keypoints from category-specific priors, whereas our modes come from the glyph itself; the only prior is a list naming each letter's moving parts, generated once for the whole alphabet by a language model. In modal kinetic typography, shape and motion are disentangled by construction: a single base outline is sculpted toward the concept, and the modal drive cannot alter it, so a letter can also be animated without being reshaped. Our method produces more articulated and smoother motion than Dynamic Typography and AniClipart at comparable or better concept alignment, with less glyph tearing than Dynamic Typography, and is preferred by human raters, including in a frozen-shape setting where motion alone must carry the concept. Our results were also preferred over Astra (GPT-6) by human raters.
1 INTRODUCTION
Automating kinetic typography requires expressive motion that remains legible and carries meaning, while prior approaches either leave points free to tear and jitter or borrow structure from category-specific priors. Modal kinetic typography instead derives motion from a glyph’s own modes, with a periodic drive and separate shape and motion representations.
1 INTRODUCTION
- The modal basis addresses two representation problems: free-form points compete between sculpting and animation, while borrowed keypoints or skeletons depend on category-specific priors.Independent point and frame updates under noisy SDS gradients can tear the outline and jitter the motion.Source: §1 · pp. 1–2
- A frozen video diffusion model supervises only the modes’ amplitudes and phases, while a once-generated per-alphabet part list determines where articulation parts begin and end.The glyph’s mechanics determine how each part bends; the part list determines its boundaries.Source: §1 · p. 2, Figure 1 · p. 2
- Glyph-derived modes and a periodic drive constrain noisy SDS updates to smooth, looping motion instead of the tearing and jitter caused by independent point-and-frame motion.The modes move neighboring outline points together, while shared whole-cycle sinusoids leave no per-frame freedom for noise to occupy.Source: §1 · pp. 2–3
- A finite-element eigensolve supplies soft whole-letter and part-localized bending modes, while closed-form rigid modes let letter parts move as blocks.The basis is computed from the glyph outline; layers are orthogonalized so part motion is not duplicated by bending modes.Source: §1 · pp. 2–3
- Shape and motion use separate representations, so the modal drive cannot alter the base outline and a letter can be animated without reshaping.This supports the frozen-shape setting, where action alone must carry the concept.Source: §1 · pp. 1, 3
2 RELATED WORK
Prior kinetic-typography work spans static concept-shaped glyphs, template-driven animation, and high-dimensional or externally structured motion. This work instead derives a vector glyph’s motion modes from its own geometry and uses a video prior to supervise their temporal coefficients.
2 RELATED WORK
- Earlier concept-driven glyph methods produce static results, while KineTy draws motion from trained templates and Differentiable Variable Fonts remain within a typeface’s defined variants.The paper’s setting instead animates vector letters under a video diffusion prior.Source: §2 · p. 3
- Dynamic Typography uses per-frame control-point displacements, while AniClipart and related methods borrow detected keypoints, skeletons, or category templates to stabilize motion.The paper contrasts these with a glyph-derived spatial basis and a periodic drive shared across frames.Source: §2 · p. 3
- Natural vibration modes are established deformation subspaces, but prior image-based approaches recover or predict motion from raster video rather than solving modes on a glyph’s vector outline.Modal warping, constrained deformation, and localized modes provide related mechanics foundations.Source: §2 · pp. 3–4
- Unlike control methods that integrate equations of motion or optimize hand-specified objectives, this method prescribes analytic temporal coefficients and lets a frozen video prior supervise amplitudes and phases.Its basis also combines alphabet-level parts, per-part rigid modes, and orthogonalized layers.Source: §2 · p. 4, Figure 2 · p. 4
3 METHOD
The method separates concept-directed shaping of a single glyph outline from its motion, which is built from whole-letter and part-level modes. A periodic drive controls those modes while the pipeline renders and optimizes the looping animation.
3 METHOD
- The pipeline separately sculpts one base outline toward the concept and drives its motion through shared coefficients, so neither parameter group can perform the other's work.The output is a looping animation whose frames are deformed vector paths.Source: §3 · p. 4
3.1 SHAPE SCULPT: GENERATING THE BASE SHAPE
- A small network sculpts the base shape under video SDS with perceptual and conformal anchors; motion is established first and discourages edits that disrupt articulation.Unlike per-frame geometry, the single template requires the concept to reside in the base shape.Source: §3.1 · pp. 4–5
3.2.1 From glyph to motion basis
- Finite-element stiffness and mass matrices from a mesh of the glyph yield whole-letter vibration modes, with the softest modes producing broad, smooth bends.The mesh is scaffolding for assembling the matrices; solved modes are interpolated back to the full-resolution outline.Source: §3.2.1 · p. 5
3.2.2 Localized modes for parts
- Localized modes confine bending to a part using stiff springs outside its region, while a shared alphabet-wide part list defines regions even for unbranched letters.Soft joint bands let limbs pivot, and finite spring stiffness allows modes to extend slightly beyond their regions.Source: §3.2.2 · p. 5
3.2.3 Hierarchical basis
- The hierarchical basis combines one whole-letter soft mode with three rigid modes and two localized bending modes per part; orthogonalization separates pivoting from bending.The basis is solved once on the source glyph and held fixed during sculpting; for three parts, it contains 16 modes.Source: §3.2.3 · pp. 5–6
3.2.4 Motion model and periodic drive
- Each mode's displacement pattern is scaled over time by a coefficient built from whole-cycle harmonics, with amplitudes learned and phases shared within each part.The drive uses three harmonics; a three-part letter has 52 animation parameters, and the shared periodic coefficients leave no per-frame freedom for jitter.Source: §3.2.4 · p. 6
3.2.5 Rendering, objective, and schedules
- Differentiable rasterization and video SDS optimize only modal amplitudes, phases, and shape-network weights, with perceptual and conformal anchors and a joint edge-length term.The drive starts first, and shape sculpting joins at step 250.Source: §3.2.5 · p. 6
4 EXPERIMENTS
The experiments compare modal animation with published baselines and Astra in full and frozen-shape single-letter tasks. They report stronger articulation against both baselines, alongside lower stretch than Dynamic Typography.
Qualitative results
- Figure 3 shows an X breakdancer, an R rooster, and an unsculpted B that jiggles while retaining its original shape.Rows show the input followed by frames at fixed scale and position, beginning with the frame closest to the input.Source: Figure 3 · p. 7
User study
- Table 1’s user study finds our method leads on both full-task questions and frozen-shape motion, although case selection favors the baselines.The study includes 44 participants for the full task and 39 for frozen shape.Source: Table 1 · p. 7, §4 · p. 8
Evaluation setup
- Each setting uses 100 generated cases across random letters and 12 typefaces, producing 16-frame animations; published baselines run at their default settings.Astra receives the caption, letter, and typeface and answers in one shot.Source: §4 · pp. 6–7
Quantitative comparison
- Our stretch is less than half Dynamic Typography’s and lower in 95% of cases, while concept similarity, temporal consistency, and motion smoothness are higher.The comparison uses a similar motion budget and favors articulated deformation over rigid translation.Source: §4 · p. 7
- 13.44 vs. 8.52 articulation when both methods move; overall, our method roughly doubles articulation while translating less, while AniClipart leads temporal consistency and stretch partly due to low-motion cases.Both methods use low-dimensional parameterizations; AniClipart also uses an as-rigid-as-possible objective.Source: §4 · p. 8
Comparison with Astra
- Astra leads several geometric and temporal metrics, but its redraw loses glyph identity; when frozen shape disallows redrawing, its articulation collapses.Its high dead time reflects both unmoving ink and samples without acceleration.Source: §4 · p. 8, Figure 4 · p. 8
5 ABLATIONS
The ablations show that free per-frame control jitters, while removing orthogonalization favors bending over pivoting. In the frozen-shape comparison, the glyph’s own modes produce fuller articulation than AniClipart or Astra.
5 ABLATIONS
- Free per-frame drive more than triples median jerk and worsens it in all 20 pairs, while temporal consistency falls.The table reports medians at step 1000 on the same 20 cases; no Articulated difference is significant.Source: §5 · p. 8, Table 2 · p. 9
- With the same frozen outline, the glyph’s modes articulate the S and H more fully than AniClipart’s keypoint rig or Astra.AniClipart holds letters closer to rest; Astra mostly declines to move the given outline.Source: Figure 5 · p. 9
- Without orthogonalization, part eigenmodes repeat rigid motion around 80%, favor bending over pivoting, and use 0.22 drive versus 0.05.Global motion also increases in 14 of 20 pairs, and the bird appears dragged rather than pivoted.Source: §5 · p. 9
- Using only whole-letter modes leaves motion concentrated at free ends: Concentration rises from 0.39 to 0.48.The removed part layer combines localization and per-part rigid motion, so this ablation does not separate their contributions.Source: §5 · p. 9
6 CONCLUSION, LIMITATIONS, AND FUTURE WORK
Modal kinetic typography builds motion from a glyph’s vibration and per-part rigid modes, with a frozen video diffusion prior supervising their amplitudes and phases. Its stated limits include periodic motion and reliance on how well the letter, typeface, and fixed part list fit the concept.
6 CONCLUSION, LIMITATIONS, AND FUTURE WORK
- The method combines a glyph’s natural vibration modes with per-part rigid modes, supervised by a frozen video diffusion prior acting on amplitudes and phases.The resulting motion is smooth and less costly to legibility than free-form optimization, while the modal drive cannot reshape the letter.Source: §6 · p. 9
- A single outline cannot change topology or turn in depth, and the periodic drive excludes actions that end somewhere other than where they start.The part list is fixed per letter, may mismatch a subject’s anatomy, covers uppercase Latin letters only, and the evaluation uses single letters rather than words.Source: §6 · p. 9
- Future work could add non-periodic motion, constraints for stationary parts, other scripts and words, and automatic letter selection for prompts.Source: §6 · p. 9
A.1 ABLATIONS, QUALITATIVELY
Figure 6 qualitatively compares the full model with ablations using matched letters, captions, seeds, and mode counts.
A.1 ABLATIONS, QUALITATIVELY
- Free per-frame drive jitters between consecutive weightlifter frames.Source: Figure 6 · p. 12
- Whole-letter modes move little beyond the hammock’s rail tops.Source: Figure 6 · p. 12
- Without orthogonalization, the bird is dragged rather than pivoted.Source: Figure 6 · p. 12
A.2 THE CAPTION CONTROLS THE MOTION
With the letter and typeface held fixed, changing the action caption changes the motion produced by the method. Dynamic Typography moves similarly for the two captions described.
A.2 THE CAPTION CONTROLS THE MOTION
- On the same K glyph, “raises his sword” produces a raised sword while “swings his sword” produces a downward sweep.Source: §A.2 · p. 12, Figure 7 · p. 12
- The letter and typeface stay fixed while the action caption varies, so the caption controls how the modes are played rather than what they are.Source: §A.2 · p. 12
- Dynamic Typography produces nearly the same action for both captions, unlike the distinct motions described for this method.Source: §A.2 · p. 12, Figure 7 · p. 12
A.3 SPEED IS A KNOB
A single fundamental frequency controls how quickly the motion plays. Integer frequency multipliers preserve seamless looping, while a half-cycle multiplier breaks it.
A.3 SPEED IS A KNOB
- Action speed is set by fundamental frequency f, which plays the basis f times per clip without changing the action.Figure 8 compares f values of 0.5, 1, and 2, producing half, one, or two lizard strides in the same frame count.Source: §A.3 · p. 12, Figure 8 · p. 13
- Integer f preserves seamless looping because every harmonic completes an integer number of cycles.At f=0.5, the loop breaks.Source: §A.3 · p. 12
- Changing f varies action rate while leaving the action unchanged.Source: §A.3 · p. 12
A.4 USER STUDY DETAILS
Participants ranked animations on motion quality and visual quality while preserving letter identity; in the frozen-shape setting, they answered only the motion question. The study brief applied equally to this method and Dynamic Typography because both deform the input outline.
A.4 USER STUDY DETAILS
- The study brief described the task as defined by Dynamic Typography, and applied equally because both methods deform the input outline.Source: §A.4 · p. 13
- Participants ranked each animation for realistic, smooth motion and for realistic, high-quality visuals that retain the letter’s identity.Each rank was assigned to exactly one animation.Source: §A.4 · p. 13
- In the frozen-shape setting, participants answered only the motion question; Figure 9 shows the randomized, anonymized presentation page.They watched each clip in full before ranking.Source: §A.4 · p. 13
A.5 IMPLEMENTATION DETAILS
The appendix details the implementation, modal basis examples, evaluation measures, and limits. Across cases, the method more often achieves substantial articulated motion while retaining legibility than the compared baselines.
A.5 IMPLEMENTATION DETAILS
- Runs use a sparse eigensolve on the source glyph before sculpting, with fixed basis, specified learning-rate schedules, seed 0, and 16-frame clips rasterized at 320 × 320.The amplitude bound is 20.7 px, with perceptual and conformal anchors weighted at 2000 each and edge length at 100.Source: §A.5 · p. 13
A.6 VISUAL EXAMPLE OF MODES
- Hand-driven examples show a whole-letter soft mode and a mode confined to one diagonal, illustrating distinct ways an eigenmode can move a glyph.Figure 10 presents both without the video prior.Source: §A.6 · p. 14, Figure 10 · p. 14
A.7 SPREAD OF ENERGY ACROSS THE BASIS
- Eigenmodes carry slightly more energy than closed-form rigid modes, while the optimizer drives both about equally hard.Source: Table 3 · p. 14
A.8 WHAT THE DRIVE CAN AND CANNOT EXPRESS
- Above the third harmonic, the full model has 0.0004 of moving power versus 0.186 for the free per-frame drive, with the former lower in all 20 cases.The table reports median moving-power shares over the 20 ablation cases and changes only the temporal parameterization.Source: §A.8 · p. 14, Table 4 · p. 15
A.9 THE ARTICULATION-PART TABLE
- The articulation-part table contains 26 entries and 75 named parts, created once for the alphabet and reused unchanged across cases, concepts, and typefaces.Figure 11 shows five letters in two of the twelve animated typefaces.Source: §A.9 · p. 15, Figure 11 · p. 15
A.10 METRICS AND LETTER FIDELITY
- The metrics distinguish rigid travel from articulation, assess letter identity, tearing, and concept similarity, and combine articulation with legibility in Moves & legible.Moves & legible requires articulated motion of at least half a stroke width while the form still reads as its letter in most frames.Source: §A.10 · pp. 15–16
A.11 EXPRESSIVENESS AGAINST LEGIBILITY
- 27% of cases both move at least half a stroke width and remain legible, versus 14% for Dynamic Typography and 12% for Astra.In the frozen-shape setting, the rates are 43% for ours, 19% for Dynamic Typography, and 1% for Astra.Source: §A.11 · p. 16, Figure 12 · p. 17
A.12 LIMITATIONS
- The fixed-outline representation cannot change topology, while the alphabet-level part partition may be too coarse for a concept and stationary parts may move unintentionally.A coarse split reduces independent articulation; typeface geometry also changes the motion. Whole-cycle drives exclude actions meant to end elsewhere, and large in-plane rotations lengthen parts rather than turning them.Source: §A.12 · pp. 16–18, Figure 13 · p. 17, Figure 16 · p. 18