acceptodds
Under review as a conference paper at ICLR 2027

Strike a Chord! Modal Kinetic Typography

Abstract

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. Across letters and typefaces, 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.

Primary area:
generative models
Keywords:
kinetic typography, text animation, vector graphics animation, video diffusion models, modal analysis, eigenmodes, physically-based animation, semantic typography

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_QytoOpenReview ↗
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

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).

Prior, driveGlobal↓Artic.↑Concen.↓Jerk↓Stretch↓Temp.↑
ModelScope, harmonic (ours)5.508.920.390.782.960.96
ModelScope, free per-frame5.499.960.362.864.650.95
p.87.65.26<.001.11.01
CogVideoX-5B, harmonic (ours)5.5714.440.381.213.420.95
CogVideoX-5B, free per-frame4.699.640.383.174.160.93
p.25.18.39<.001.67.06

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).

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!

Reviewer_grqkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a structured representation for semantic vector-glyph animation. The method constructs a low-dimensional motion basis from whole-glyph and localized elastic eigenmodes, augmented with per-part rigid translation and rotation modes. Motion over time is parameterized using a small number of periodic harmonics, whose amplitudes and phases are optimized under a frozen text-to-video diffusion prior via SDS. A separate shape-sculpting branch modifies a shared base glyph, aiming to disentangle shape from motion.

The paper evaluates both a full shape+motion setting and a frozen-shape setting, comparing against Dynamic Typography, AniClipart, and Astra. The reported results show improved articulated motion, reduced jitter/tearing, and favorable human rankings.

Critical strengths
  1. The low-dimensional periodic motion parameterization is well motivated and reasonably well validated.

    In particular, the temporal ablation is convincing: replacing the harmonic drive with independent per-frame coefficients substantially increases jerk (0.78 to 2.86 at the median), with worse jerk in all 20 paired cases. This provides clear evidence that the proposed temporal parameterization is effective at suppressing high-frequency jitter under SDS.

Critical weaknesses
  1. The experiments do not sufficiently isolate the benefit of the mechanics-derived spatial basis.

    The temporal ablation convincingly demonstrates the benefit of the harmonic drive, but there is no comparably controlled experiment establishing that elastic eigenmodes are preferable to a simpler low-dimensional smooth deformation basis. The "global modes only" ablation simultaneously removes part localization and per-part rigid motion, while the other spatial ablations either retain eigenmodes or only vary their number/orthogonalization. In fact, Table 2 reports no statistically significant difference in articulated motion for any of these variants.

    Therefore, it remains unclear how much of the reported improvement should be attributed specifically to the proposed vibration modes, rather than to more generic factors such as dimensionality reduction, part decomposition, and temporal regularization. Since the glyph-derived modal basis is presented as a central contribution of the paper, I find this missing control important.

  2. The evaluation is substantially stronger at measuring motion regularity than semantic correctness of motion.

    A central motivation of the paper is that the motion itself should convey the prompted action, particularly in the frozen-shape setting. However, most quantitative metrics measure deformation magnitude, smoothness, tearing, temporal consistency, and glyph legibility rather than whether the requested action is actually performed.

    The paper itself notes that shuffling video frames changes the X-CLIP score by less than 2%, suggesting that the Concept metric is largely insensitive to the temporal semantics of the action. The user study also asks participants to rank motion realism/smoothness and visual quality, rather than directly asking which animation best matches the action described in the caption. Apart from a small qualitative example with different action captions, there is little systematic evidence that the proposed method improves action-level text-motion alignment.

    As a result, the paper more convincingly demonstrates that the glyphs move more smoothly and articulately while remaining legible than that the motion itself more faithfully expresses the requested semantic action.

Questions for authors
  1. Under the same harmonic drive, part partition, loss functions, optimization schedule, and basis dimensionality, how does the proposed elastic eigenmode basis compare against a simpler smooth low-dimensional deformation basis (or a part-rigid-only basis)? In particular, does the mechanics-derived basis provide a significant improvement in the articulation-legibility tradeoff or in tearing?

  2. Can the authors provide a systematic evaluation of action-level semantic correctness on randomly sampled, unfiltered test cases? For example, a human study could directly ask which animation best matches the action described by the caption, rather than primarily evaluating smoothness/realism. Such evidence would materially affect my assessment of the paper's central claim that motion itself carries the concept.

3 comments

Thank you for your valuable feedback on our work. Below please find our response to your questions and requests.

Basis control at the same drive, parts, losses, schedule and DOF

Thank you for the request. As Reviewer kQTR raised a similar point, please also refer to our full response to Reviewer kQTR for the full setup and results. In short, against generic smooth fields with the same parts and DOF, the articulation-legibility trade-off itself does not change significantly (Articulated 8.68 vs. 9.67, p = 0.15; Reads 0.69 vs. 0.75, p = 0.78). Actual tears are equally rare with both bases (the moving letter keeps its structure in 0.78 vs. 0.76 of frames, p = 0.65), but generic fields bring the strokes much closer to tearing: they reach the same motion by stretching the strokes, with 1.7 times the strain, the thinnest stroke narrowing to 0.65 of its rest width against 0.79, and a stroke pinching below half its width in 36% of examples against 16% (all p ≤ 0.005). Part-rigid modes alone sit at the other end of the trade-off: the letter reads more often (0.84 vs. 0.69) but articulates 30% less (6.05 vs. 8.68).

Action-prompt alignment

We note that this request echoes a similar one from Reviewer kQTR. We agree that our user studies and experiments focused more on the concept alignment and smoother motion, as opposed to action-prompt alignment. Having said that, one can find implicit claims on action alignment at several places in the paper, e.g., the appendixes and Fig. 1 caption.

To address this we have conducted a new user study at the reviewer's request. In this study, we follow the same protocol as in the paper for selecting the examples, with randomization to remove bias. Please refer to line 411 in the paper, where we should have stated "randomly sampled" (5 out of 20) rather than just "sampled." For the new study, we randomly selected 10 out of 40 using similar methodology.

In the new study, the participants are explicitly asked to rank results based on how well the actions in the animations follow the prompt. We color-coded the texts related to actions in each question to make sure this emphasis is not missed. Please see the study setup and questions [Supplementary-Rebuttal HTML → User Study].

With 20 responses so far, the results are (average points per page, with 2 for first, 1 for second and 0 for last place):

MethodSection 1 (Full Task)Section 2 (Frozen Shape)
Ours1.481.46
Dynamic Typography0.92–
AniClipart–0.95
Astra0.610.60 (font only)

We will report the final numbers when the study closes.

Although the cases, participants and questions differ from the study in the paper, our margin over the main baseline is, if anything, larger when participants judge the action: 0.59 points over Dynamic Typography (0.43 in the paper) and 0.52 over AniClipart (0.31 in the paper). We are happy to include the new study in the revision. While the results do support that our method outperforms the baselines (Dynamic Typography, AniClipart and Astra) on action-prompt alignment, we will still go over all claims about this in the paper to make sure that we do not make any unsubstantiated claims.

AuthorsOpenReview ↗

Any feedback?

Dear Reviewer grqk:

We understand that this is a busy few days for many. If you have any feedback which may require us to perform additional tests, please let us know soon so we could find time to work on them. We believe that the last chance we have to respond to such requests is only two days away. Thanks again for your initial review and the questions.

Sorry for pressing again on this, but we have less than a day to possibly address any further questions. The action-prompt alignment question was a good one and we now have a user study to address it. Is it sufficient? We would really appreciate to get some feedback before we lose the voice to further elaborate ...

Reviewer_kQTROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents a novel framework for kinetic typography that animates vector glyphs to express semantic concepts using glyph-derived vibration modes. By solving a finite-element eigenproblem on an interior mesh derived from the glyph outline, the method obtains whole-letter and localized bending modes, which are combined with closed-form per-part rigid-motion modes. Rather than learning unrestricted per-frame boundary displacements, it uses a frozen video diffusion model to supervise a low-dimensional modal motion representation through score distillation sampling (SDS). The modal coefficients are parameterized by a few integer-harmonic sinusoids, whose amplitudes and phases are optimized to produce smooth, seamlessly looping motion. A separate MLP sculpts a single base outline, while disabling this stage enables animations that convey concepts through movement alone. Experiments demonstrate improved articulation, lower jerk, higher user preference, and a better trade-off between motion magnitude and legibility relative to Dynamic Typography and AniClipart in their respective evaluation settings.

Critical strengths
  1. The authors derive a motion basis from the glyph’s own geometry by treating it as a thin elastic sheet and computing its low-energy deformation modes. Combining these bending modes with part-level rigid translations and rotations allows a frozen video diffusion model to supervise only the amplitudes and phases of the modal motion. While vibration modes themselves are classical, their integration into a part-aware, diffusion-guided framework for semantic kinetic typography is a novel and well-motivated contribution.

  2. Since unrestricted per-frame motion can lead to torn outlines and temporal jitter, the method constrains the degrees of freedom through spatially smooth modes and a periodic harmonic drive. The ablation with independently learned per-frame coefficients provides convincing evidence that the periodic parameterization reduces temporal jitter.

  3. Minimal motion can trivially yield high smoothness and shape-preservation scores. Recognizing this issue, the authors jointly evaluate whether the generated animations exhibit substantial articulated motion while remaining legible. Furthermore, they evaluate a frozen-shape setting in which base-shape sculpting is disabled, testing whether the prompted concept can be expressed through motion alone.

Critical weaknesses
  1. While the authors emphasize generating semantics-driven animations, direct quantitative evidence that the motion accurately reflects the prompted action remains limited. As noted in Appendix A.12, shuffling video frames changes the X-CLIP score by less than 2%, and the authors accordingly treat it as a concept-similarity measure rather than a motion metric. However, the user study primarily evaluates naturalness, smoothness, and visual quality, without separately measuring action–prompt alignment. The core claim would be strengthened by a controlled human evaluation of action-recognition or motion–prompt matching accuracy, using different action prompts while keeping the glyph and subject fixed.

  2. Although the ablation study on temporal representation is persuasive, the incremental contribution of the finite-element vibration modes themselves remains insufficiently isolated. The paper lacks controlled comparisons against simpler low-dimensional alternatives, such as using only part-level rigid modes or other smooth geometric bases. Comparisons using the same temporal drive and optimization settings would help determine whether the performance gains stem specifically from the physics-inspired modal basis or primarily from part segmentation and restricted degrees of freedom.

Questions for authors
  1. Could the authors conduct a controlled human evaluation of action–prompt alignment, using different action prompts while keeping the glyph and subject fixed, to verify whether the generated motion conveys the intended action?

  2. Could the authors compare the proposed modal basis against part-level rigid modes alone and other smooth low-dimensional bases, using the same temporal drive and optimization settings, to isolate the contribution of the finite-element vibration modes?

5 comments

Thank you for your positive feedback and recommendation. Below we address your questions and requests.

Basis control (part-rigid only and another smooth basis)

Thank you for raising this point. We added Appendix A.9, which isolates the contribution of the FEM modes with three controls. Each control keeps the drive, losses and schedule fixed:

  1. Generic smooth fields: every mode is replaced with the lowest cosines over the letter's bounding box. The parts and the number of modes per part stay the same, so the DOFs are identical and only the basis differs.
  2. Part-rigid only: the eigenmodes are removed, which shows what the parts alone provide.
  3. Without parts: the whole letter is one part, and the glyph's FEM modes are compared with the vibration modes of its bounding box, which ignore the letter's shape.

(1) addresses the restricted DOFs, and (3) addresses the part segmentation. All controls use 50 random examples after 1,000 steps, with paired Wilcoxon tests. Because the basis also shapes the sculpting, deformation is measured on the original letter moved by each basis's motion, so every basis is compared on the same geometry. Results are in Table 3 (for (1) and (2)), the text (for (3)) and Fig. 11. The Supplementary-Rebuttal HTML (Basis Comparison tab) shows all three side by side, and the animations make the difference much clearer than the numbers.

(1) Generic smooth fields. Both bases move the letter equally (0.183 vs. 0.184), with no significant difference in articulation, concept alignment or legibility. The eigenmodes bend and turn the parts (bend share 0.27 vs. 0.08, higher in 49/50), while generic fields stretch them (stretch share 0.79 vs. 0.45, higher in 50/50). As a result, generic fields strain the letter 1.7× as much, thin the thinnest stroke to 0.65 of its rest width vs. 0.79, and pinch a stroke below half its width in 36% vs. 16% of examples. The eigenmode motion is also smoother (jerk 0.89 vs. 1.03). All p ≤ 0.03.

(2) Part-rigid only. Parts bend much less (0.12 vs. 0.27), articulation drops from 8.7 to 6.1, and motion drops from 0.183 to 0.148 (p ≤ 0.02). In exchange, the letter deforms least (strain 0.064 vs. 0.110) and reads as its letter more often (0.84 vs. 0.69, p = 0.01), with slightly higher concept alignment (19.69 vs. 19.30, p = 0.049).

(3) Without parts. On the same 50 examples, there is no significant difference in articulation (10.1 vs. 10.5, p = 0.052), jerk or legibility. The box modes stretch the letter: strain per unit motion is 1.79 vs. 1.02 (higher in 41/50) and bend share is 0.08 vs. 0.30 (lower in 49/50), both p < 0.001. The thinnest stroke narrows to 0.69 of rest width vs. 0.91 (p < 0.001). With no parts, the advantage comes from solving the modes on the glyph's own shape.

Summary. At a fixed part structure and parameter count, the eigenmodes move the letter as much as generic fields, but they do it by bending parts rather than stretching them. Strokes keep more of their width and pinch less than half as often. Removing the eigenmodes keeps the letter intact but cuts articulation by ~30%. The advantage also holds against bounding-box modes without parts, so it comes from the glyph's own shape, not from the part segmentation or the restricted DOFs.

Table 3. Basis controls (medians; Pinch = % of examples with a stroke below half its width).

BasisArtic.↑Concept↑Jerk↓Motion↑Reads↑Turn↑Bend↑Stretch↓Strain↓Thin↑Pinch↓
vs. generic
FEM (ours)8.6819.300.890.1830.690.260.270.450.1100.7916%
Generic9.6719.131.030.1840.750.120.080.790.1900.6536%
p.146.716.029.426.782<.001<.001<.001<.001.003.005
vs. part-rigid
FEM (ours)8.6819.300.890.1830.690.260.270.450.1100.7916%
Part-rigid6.0519.690.920.1480.840.540.120.350.0640.918%
p<.001.049.774.020.010<.001<.001<.001<.001<.001.024

Artic., Concept, Jerk: on output clips; others: on the original letter moved by each basis. Bold: better of the pair, p < 0.05; italic: better, n.s.

Action recognition with glyph and subject fixed

Note that Reviewer grqk raised a related point. Please see our response there, where we describe a new user study to specifically evaluate action-prompt alignment when compared to baselines.

This test is more demanding than the user study: the glyph is fixed and only the prompt varies. Appendix A.4 (Fig. 7) already shows one such case, where ours beats Dynamic Typography. We extended it in both settings of the paper. In the full task, the subject is also fixed and only the action varies; because the caption also reshapes the glyph, clips can differ by shape. The frozen-shape setting isolates motion, so there we give each letter a variety of motion prompts. We manually labeled each output as following, partly following, or not following its prompt (all clips and labels: rebuttal HTML, Multi-Prompt Comparison).

  • Full task, 5 letter-subject pairs x 3 actions, vs. Dynamic Typography on the same letters, typefaces and prompts: ours followed / partly / missed = 10 / 4 / 1, against 5 / 7 / 3. All three of our motions matched in 1 of 5 pairs; Dynamic Typography's in none.
  • Frozen shape, 5 letters x 3 varied motion prompts, vs. AniClipart: 9 / 5 / 1, against 1 / 8 / 6. AniClipart missed all three prompts for the M: its as-rigid-as-possible and bone-length losses keep its motion small, and here the action must come from motion alone.
  • Shared failures: of the 11 outputs where ours falls short, the baseline also falls short on 10 (e.g., an M squeezing like an accordion fails for both).

Why the results are mixed: Success depends on how well the letter and typeface suit both the subject and the action, especially with the shape frozen. Appendix A.15 (Fig. 16) shows the converse: with the caption fixed, the same letter animates differently across typefaces. With the glyph fixed, every prompt must suit one shape, so finding several suitable prompts is hard (the prompts were written by an LLM). Two limitations from A.15 also apply to any method that deforms the input glyph: the drive is periodic, so actions that end in a different pose cannot be expressed, and large rotations are limited.

Our claim of better action alignment rests primarily on new user study (response to Reviewer grqk). In the revision, we will report these results and discuss the test as a limitation

AuthorsOpenReview ↗

Any feedback?

Dear Reviewer kQTR:

We understand that this is a busy few days for many. If you have any feedback which may require us to perform additional tests, please let us know soon so we could find time to work on them. We believe that the last chance we have to respond to such requests is only two days away. Thanks again for your initial review and the questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.