- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper presents a method for animating vector letters from text prompts while maintaining legibility. Motion is represented using the glyph’s finite-element vibration modes, both global and part-localized, together with closed-form rigid modes for individual parts. Each mode’s coefficient is driven by a few harmonics that complete whole cycles within the animation, producing smooth, seamless loops. A frozen video diffusion model guides the optimization of motion amplitudes and phases through SDS. In the full model, the same guidance also trains a separate network to refine a single base outline, keeping shape and motion separately parameterized. The method produces more articulated motion than Dynamic Typography and AniClipart, reduces jitter and stretching relative to Dynamic Typography, and is preferred by participants in user studies. Ablation results further show that the periodic driving functions play a central role in reducing jitter.
Critical strengths
The method combines the glyph’s own vibration modes with whole-cycle harmonics to constrain motion in both space and time, giving SDS a compact set of parameters to optimize. The temporal parameterization guarantees smooth coefficient curves and seamless looping. Experiments show less jitter and boundary stretching than Dynamic Typography, while ablations support the periodic drive’s role in reducing jitter.
In the frozen-shape setting, where static shape sculpting is disabled, the method achieves 43% M&L, compared with 19% for AniClipart and 1% for Astra. The result shows that the method satisfies both requirements more often without relying on reshaping the base glyph.
Critical weaknesses
Vibration modes have long been used in animation. For example, Benchekroun et al. (2024) combined elastic modes with sine waves to drive motion in soft bodies. The authors apply these ideas to animating vector letters, using video SDS to guide the motion of the whole letter and its individual parts. This design has practical value, but the paper still lacks a direct comparison: keep the original part divisions and periodic drive, and replace the FEM modes with another smooth spatial basis using a similar number of parameters. Without such a comparison, it remains difficult to judge how much additional benefit the FEM modes themselves provide.
In addition, the paper uses several custom metrics, but there has been limited validation of how well these metrics reflect the perceptual qualities they aim to measure. For example, Articulated measures non-rigid deformation and does not directly measure semantically correct part motion. The 88% recognition accuracy reported for Reads-as-letter was validated on undeformed glyphs. This does not establish that it is equally reliable on glyphs deformed by the animation. The authors also note that shuffling the frames changes Concept by less than 2%, indicating that the score is not very sensitive to the order of the motion. The study instructions mention semantic fidelity, but participants did not score action matching separately. Along with the qualitative example in Appendix A.4, the results suggest that the method can respond to action prompts, though its advantage in action matching still lacks direct quantitative validation.
Questions for authors
In Table 1, “Conc.” refers to Concept, while in Table 2 it refers to Concentration. Using the same abbreviation for two different metrics is confusing, especially since their preferred directions differ. Could the authors use distinct labels or spell out the names in the table captions?
The experiments use a single video prior. With a stronger video prior, would the free per-frame drive still produce much more jitter than the periodic drive? Do the authors have evidence from another prior to help assess this?
The part table is generated once by an LLM and then fixed. How much would the results change with another reasonable table?
3 comments
Thank you to the author for addressing my confusion and concerns. I keep the positive score.
Thank you again, Reviewer Qyto, for your insightful questions!
We are grateful to the reviewer for their positive feedback and recommendation on our work. Below please find our response to your requests.
Replace FEM modes with another smooth basis, same parts + drive
Thank you for the request. As Reviewer kQTR raised a similar point, please refer to our full response to Reviewer kQTR.
Conc. issue
Fixed: Table 1 now reads "Concept" and Table 2 "Concen."
Action matching not scored
Thanks for the question, which Reviewers kQTR and grqk both alluded to as well. Please refer to our response to Reviewer grqk with the new user study that explicitly tested action-prompt alignment, and our response to Reviewer kQTR on a request to test this by fixing the glyph and subject.
Metric validation
Reads-as-letter was validated only on undeformed glyphs, so we now also test it on deformed ones. Warping clean glyphs by smooth non-rigid fields of known size d (measured as in Articulated), the matcher stays close to its undeformed accuracy at the deformation level of most of our frames (80% at d = 10% of glyph size, where our median frame has d = 8.9%, against 88% undeformed) and degrades gradually beyond it (45% at d = 30%). Its errors grow with deformation, so if anything it is conservative for methods that move more, such as ours.
Articulated measures deformation beyond rigid motion, not whether the right parts move, so on its own it cannot separate correct motion from merely articulated motion; we will state this in its definition. It does, however, agree with human judgement: in the paper's user study, the clip with more articulation was ranked higher on Motion in 49 of 60 within-example pairs (82%, p < 0.001), and in 16 of 20 pairs without our method.
Concept barely depends on frame order (under 2% change when the frames are shuffled), so it measures whether the concept appears rather than whether the action is performed. Action matching is now scored separately by the new user study (see our response to Reviewer grqk).
Evidence from a second video prior
We repeat the drive ablation of Table 2 with CogVideoX-5B (Yang et al., 2024), a larger and more recent video model, in place of ModelScope as the SDS prior, on the same 20 cases with the same schedule, timestep range, guidance scale and weighting. Under both priors the harmonic drive has much lower Jerk than the free per-frame drive in all 20 cases: 0.78 vs. 2.86 under ModelScope and 1.21 vs. 3.17 under CogVideoX-5B (both p < 0.001). The smoothness therefore comes from the drive, not from a particular prior. The harmonic drive's higher Jerk under CogVideoX-5B (1.21 vs. 0.78) reflects a preference of that prior for faster motion, not jitter: it puts a median 76% of the motion's energy in the highest allowed harmonic (8% under ModelScope), and a pure sinusoid at that frequency alone has Jerk 1.23. No other metric differs significantly under CogVideoX-5B (all p ≥ 0.06).
Medians over the 20 cases; p from paired Wilcoxon tests. Bold: better and significant (p < 0.05); italic: better but not significant.
Sensitivity to Part Table
Both part tables were written once for the whole alphabet by a language model, but from different questions. The paper's table starts from how each letter is written (its pen strokes) and splits the strokes that hide a clear moving part (e.g., an L into an upright and a foot). For table B, a fresh session with no access to the first table was asked only for each letter's parts if it were animated as a character or object.
We ran the full model with each table on the 20 ablation cases (rerun from scratch on a new machine for this comparison), with everything else unchanged. No metric differs significantly (paper table vs. table B: Articulated 7.58 vs. 9.86, Jerk 0.69 vs. 0.81, Stretch 4.04 vs. 3.36; all p ≥ 0.08), either on all 20 cases or on the 7 letters whose parts changed. The only change is where the letter bends, which is the table's job; how much it moves, how smoothly, and whether it reads stay the same. However, as we also note in our limitations (Appendix A.15), the parts do affect the motion: too coarse a division takes articulation away, as when the starfish's four arms are merged into one part (Figure 17).