Research map / Video / Video generation
Past as State, Present as Attention: Persistent-State Blockwise Flow Matching for Long-Horizon Co-Speech Motion Generation
Yi Liu, Xiangyue Zhang, Jia Ma, Jianfang Li, Yisheng He, Jianqiang Ren
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a persistent-state blockwise flow matching model for long-horizon co-speech motion generation, carrying a compact persistent state across segments instead of short seeds or ever-longer explicit histories.
Problem
Long-horizon co-speech motion generation suffers from temporal drift when conditioning on short seed frames, and from high cost and poor learnability when retaining long explicit motion histories.
Approach
Separates the past as a persistent state carried across blocks from the present handled by attention within each block, generating motion blockwise with flow matching; the abstract only partially describes this.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add persistent-state history (State past, Attn current) instead of no history (ablation) | PASPA · BEAT2 1 Speaker | FGD: 7.065 -> 5.216 | Improved |
| Use persistent state for past instead of full token-attention history (ablation) | PASPA · BEAT2 1 Speaker | FGD: 5.710 -> 5.216 | Improved |
| Use attention for current block instead of state-based modeling (ablation) | PASPA · BEAT2 1 Speaker | FGD: 7.652 -> 5.216 | Improved |
| Switch from single-block to multi-block supervision naively (ablation) | PASPA · BEAT2 1 Speaker | FGD: 7.410 -> 8.456 | Hurt |
| Add weakly monotonic timesteps (ablation) | PASPA multi-block · BEAT2 1 Speaker | FGD: 8.456 -> 6.346 | Improved |
| Add adaptive loss weighting (ablation) | PASPA multi-block · BEAT2 1 Speaker | FGD: 6.346 -> 5.614 | Improved |
| Add stop-gradient (SG) (ablation) | PASPA · BEAT2 1 Speaker | FGD: 5.614 -> 5.216 | Improved |
| Add logit-normal time sampling (ablation) | PASPA · BEAT2 1 Speaker | FGD: 5.216 -> 4.450 | Improved |
| Add condition dropout during training (ablation) | PASPA · BEAT2 1 Speaker | FGD: 4.450 -> 3.937 | Improved |
| Add history dropout without guidance (ablation) | PASPA · BEAT2 1 Speaker | FGD | Hurt |
| Use full Hybrid CFG instead of condition-only CFG (ablation) | PASPA · BEAT2 1 Speaker | FGD: 3.883 -> 3.532 | Improved |
| Add prefill-decode separation (ablation) | PASPA · synthetic ultra-long rollouts | latency | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SemTalk — SemTalk: Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis (ICCV 2025) · better on 1 of 1 shared result
- GestureLSM — GestureLSM: Latent Shortcut based Co-Speech Gesture Generation with Spatial-Temporal Modeling (ICCV 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Experimental setup
Builds on: flow matching
Closest papers by meaning
- Mitigating Error Accumulation in Co-Speech Motion Generation via Global Rotation Diffusion and Multi-Level Constraints (AAAI 2026 · similarity 0.89)
- Long-Term TalkingFace Generation via Motion-Prior Conditional Diffusion Model (ICML 2025 · similarity 0.87)
- SemTalk: Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis (ICCV 2025 · similarity 0.87)
- Long-Form Speech Generation with Spoken Language Models (ICML 2025 · similarity 0.86)
- TANGO: Co-Speech Gesture Video Reenactment with Hierarchical Audio Motion Embedding and Diffusion Interpolation (ICLR 2025 · similarity 0.85)
- Deep Compositional Phase Diffusion for Long Motion Sequence Generation (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
