A computational evaluation of speech-synchronized, full-body gesture generation reported lower spatial-control errors for latent-guidance schedules than for the listed baseline methods. In the same BEAT2 and GestureLSM benchmark, Progressive Chunk Guidance reported higher errors on the spatial-control measures than Synchronous Chunk Guidance, although the two schedules had different values on other reported motion metrics.
Steering a generator during production
InteractGesture applies spatial guidance at inference time, while the generator is producing motion. It treats the generator as a black box and backpropagates spatial-control gradients into estimates of the target latent variables, the internal numerical representations used during generation, while leaving the model parameters untouched.
Its Progressive Chunk Guidance schedule divides generation into chunks with staggered delays. It refreshes context from intermediate target estimates and keeps overlapping chunks editable, so spatial control can be carried across chunk boundaries while the process remains compatible with streaming generation.
A bounded benchmark
The reported test used BEAT2 sequences with the pretrained GestureLSM co-speech motion generator. The generator weights and RVQ-VAE body-part decoders were frozen. Motion was synthesized at 30 frames per second in 128-frame pose chunks.
The benchmark tested one, two and five keyframe targets per chunk across five joint configurations, giving 15 control settings. It constructed three-dimensional target positions directly from ground-truth SMPL-X poses and aggregated scores over the controlled joints and the frames in which constraints were active.
The comparison reported average control error, trajectory error, location error, Fréchet Gesture Distance, beat consistency and diversity. These measures covered target-following performance alongside other reported properties of the generated motion.
Lower reported errors than listed baselines
The listed baseline results included an Inverse Kinematic Fréchet Gesture Distance of 1.043 and an average error of 19.973 centimetres. Chunk-Wise ControlNet had a Fréchet Gesture Distance of 0.491, a normalized location error of 0.976 and an average error of 46.159 centimetres. Sequential Chunk Guidance had an average error of 11.701 centimetres.
The latent-guidance schedules had lower reported average errors than the listed geometric or feed-forward baselines in this comparison. The reported comparisons did not include confidence intervals, variance estimates or statistical tests.
Progressive and synchronous schedules show different results
Synchronous Chunk Guidance reported an average error of 4.669 centimetres, a normalized trajectory error of 0.092 and a normalized location error of 0.100. Progressive Chunk Guidance reported an average error of 6.335 centimetres, a normalized trajectory error of 0.267 and a normalized location error of 0.161.
The schedules also differed on the other reported measures. Synchronous guidance recorded a Fréchet Gesture Distance of 0.442, beat consistency of 0.750 and diversity of 12.234. Progressive guidance recorded 0.431, 0.762 and 11.760, respectively. Thus, the Progressive schedule had lower reported Fréchet Gesture Distance and diversity, and higher reported beat consistency, while its spatial-control errors were higher in this benchmark.
A component ablation compared the full guidance process with sampling-only, post-refinement-only and uncontrolled configurations. At Delay 0, average error was 4.669 centimetres with Full Guidance, compared with 11.471 centimetres for sampling only, 9.428 centimetres for post-refinement only and 59.308 centimetres without control. At Delay 1, the corresponding values were 6.335, 12.740, 12.361 and 64.984 centimetres. Full Guidance had the lowest reported average error at both delays.
What this benchmark cannot establish
The numerical results come from a computational evaluation whose number of BEAT2 sequences and generated samples was not reported. The analysis also provided no confidence intervals, variance estimates, hypothesis tests or sample-level uncertainty estimates.
The reported evidence is limited to BEAT2 and GestureLSM, so it does not establish performance across other generators or datasets. No direct human perceptual test or computational latency benchmark was reported, leaving unanswered whether the metric differences correspond to human judgments of naturalness and speech synchrony or to practical real-time limits.
The results support a bounded conclusion: inference-time spatial guidance was evaluated in a BEAT2 and GestureLSM setup, and Progressive Chunk Guidance was described as compatible with streaming while reporting higher spatial-control errors than Synchronous Chunk Guidance in this comparison. The evidence does not establish model-agnostic performance, human-perceived naturalness or real-time feasibility.
Paper data and sources
Original title: InteractGesture: Progressive Chunk Guidance for Continuous Streaming Co-Speech Gesture Control
Authors: Ekkasit Pinyoanuntapong, Ajinkya Deogade, Paul Streli et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text