Skip to content

InterEdit: Navigating Text-Guided Multi-Human 3D Motion Editing

Conference: CVPR 2026 arXiv: 2603.13082 Code: GitHub Area: 3D Motion Generation / Diffusion Models Keywords: multi-human motion editing, text-guided, interaction-aware, frequency-domain alignment, conditional diffusion models

TL;DR

This paper proposes InterEdit, the first text-guided multi-human 3D motion editing framework. Through two alignment mechanisms—Semantic-Aware Plan Token Alignment and Interaction-Aware Frequency Token Alignment—InterEdit achieves precise editing of two-person interactive motions within a conditional diffusion model, while preserving source motion consistency and interaction coherence.

Background & Motivation

Text-guided single-person motion editing has achieved notable progress, yet extending it to multi-human scenarios introduces unique challenges:

Scarce paired data: No multi-human motion editing datasets exist in the form of (source motion, target motion, editing instruction) triplets.

Complex interaction semantics: Motion meaning arises not only from individual actions but also from spatiotemporal coupling—synchronization, phase alignment, role switching, and contact timing.

Stricter editing constraints: Edits must "change the requested part and preserve the rest," yet in interactive settings even minor temporal offsets can alter semantics.

Core gap: Existing single-person editing methods (MotionFix, MotionLab) ignore interaction coupling, and naively concatenating dual-person features disrupts coordination; multi-human generation methods (InterGen, TIMotion) lack mechanisms for "what to change vs. what to preserve." No dedicated multi-human motion editing benchmark currently exists.

Method

Overall Architecture

InterEdit is a conditional diffusion framework parameterized with Start_X, directly predicting clean motion \(\hat{\mathbf{x}}_0 = \mathcal{D}_\theta(\mathbf{x}_t, t; \mathbf{c}_{\text{text}}, \mathbf{c}_{\text{src}})\). The backbone is a Transformer-based denoiser with conditions injected via AdaLN:

\[\mathbf{e}_t = \mathrm{EmbedTime}(t) + W_{\text{text}}\mathbf{c}_{\text{text}} + W_{\text{src}}\mathbf{c}_{\text{src}}\]

The core contributions lie in two auxiliary alignment mechanisms: Semantic-Aware Plan Token Alignment and Interaction-Aware Frequency Token Alignment.

Key Designs

  1. Symmetric Interleaved Token Aggregation (base architecture)

A causal interleaved sequence is constructed to model bidirectional temporal influence and role switching between persons A and B. Given motion tokens \(\mathbf{x}_c^A, \mathbf{x}_c^B \in \mathbb{R}^{L \times C}\), an interleaved sequence \(\mathbf{x}_{\mathrm{cii}}\) and a role-swapped symmetric sequence \(\mathbf{x}_{\mathrm{sym}}\) are built:

\(\mathbf{x}_{\mathrm{cii}}(2\ell-1) = \mathbf{x}_c^A(\ell), \quad \mathbf{x}_{\mathrm{cii}}(2\ell) = \mathbf{x}_c^B(\ell)\)

After concatenation, a Transformer processes the sequence; de-interleaving and role-perspective fusion yield global features, complemented by an LPA (Localized Pattern Amplification) branch for short-range temporal patterns.

  1. Semantic-Aware Plan Token Alignment (semantic guidance)

\(N_M=16\) learnable Plan Tokens \(\mathbf{P} \in \mathbb{R}^{N_M \times 2C}\) are appended to the denoiser sequence. At Transformer block \(L_p\), they are projected into semantic space and aligned with target motion embeddings \(\mathbf{z}_{\text{tgt}} = f_T(\mathbf{x}_0)\) extracted by a frozen motion teacher encoder:

\(\mathcal{L}_{\text{plan}} = \frac{1}{N_M}\sum_{k=1}^{N_M}\left[-\log\frac{\exp((\tilde{\mathbf{z}}^{(k)})^\top \tilde{\mathbf{z}}_{\text{tgt}} / \tau)}{\sum_n \exp((\tilde{\mathbf{z}}^{(k)})^\top \tilde{\mathbf{z}}_{\text{tgt}}^{(n)} / \tau)}\right]\)

Aligned via InfoNCE loss, the Plan Tokens provide high-level editing semantic guidance to motion tokens through self-attention.

  1. Interaction-Aware Frequency Token Alignment (interaction dynamics)

Two interaction signals are constructed: mean \(\mathbf{z}_S = (\mathbf{x}^A + \mathbf{x}^B)/2\) (synchronization component) and difference \(\mathbf{z}_D = \mathbf{x}^A - \mathbf{x}^B\) (opposition component). DCT is applied to each, and frequency-band energy descriptors are obtained by pooling over low/mid/high bands, yielding six band-energy descriptors:

\(\mathbf{E}(\mathbf{C};b) = \sqrt{\frac{1}{|b|}\sum_{k \in b} \mathbf{C}[k]^2 + \epsilon}\)

The band energies are projected into six Frequency Tokens and injected into the sequence; at block \(L_f\) they are decoded and aligned to the target motion's band energies via a weighted regression loss: \(\mathcal{L}_{\text{freq}} = \frac{1}{N_f}\sum_i w_i \|\hat{\mathbf{g}}_i - \mathbf{g}_i(\mathbf{x}_0)\|_2^2\). High-frequency terms are down-weighted by 0.25 during training, and frequency tokens are randomly dropped with probability \(p_f=0.04\) to prevent overfitting.

Loss & Training

Total objective: \(\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{motion}} + \lambda_p \mathcal{L}_{\text{plan}} + \lambda_f \mathcal{L}_{\text{freq}}\)

The motion loss comprises: diffusion reconstruction \(\mathcal{L}_{\text{diff}}\), velocity \(\mathcal{L}_{\text{vel}}\), foot contact \(\mathcal{L}_{\text{foot}}\), bone length \(\mathcal{L}_{\text{BL}}\), plus interaction losses (distance map \(\mathcal{L}_{\text{DM}}\), relative orientation \(\mathcal{L}_{\text{RO}}\)). \(\lambda_p = 0.03\), \(\lambda_f = 0.01\). Inference uses Synchronized CFG (\(\gamma=3.5\)), jointly dropping text and source motion conditions. Sampling is performed with DDIM in 50 steps. The model has 358.8M parameters (85.0M trainable), trained for 1500 epochs on 8× RTX Pro 6000 GPUs.

Key Experimental Results

Main Results

Evaluation on the InterEdit3D test set (5,161 triplets, 80/10/10 split):

Method FID↓ g2s R@1↑ g2s R@3↑ g2t R@1↑ g2t R@3↑
MotionFix 2.547 2.51 6.76 3.86 7.73
MotionLab 0.550 7.90 16.43 13.26 20.69
InterGen 0.624 9.52 18.91 18.93 31.64
TIMotion 0.445 12.54 22.33 24.97 40.68
InterEdit 0.371 17.08 29.32 30.82 47.65

Compared to the strongest baseline TIMotion: g2t R@1/2/3 improve by +5.85/+7.07/+6.97 respectively, and FID decreases by 16.7%.

Ablation Study

Configuration FID↓ g2t R@1↑ g2t R@3↑
w/o plan + freq 0.445 24.97 40.68
only plan token 0.367 28.72 43.50
only freq token 0.380 28.75 44.05
plan + freq (full) 0.371 30.82 47.65

Ablation over frequency token dropout rate shows \(p_f=0.04\) to be optimal, balancing overfitting and signal strength.

Key Findings

  • Multi-human generation baselines (InterGen/TIMotion) substantially outperform single-person editing baselines, confirming the necessity of interaction modeling.
  • Plan Tokens and Frequency Tokens serve complementary roles: the former guides "what to change," the latter stabilizes "how to change it."
  • Combining both yields larger gains than either alone (g2t R@3: 40.68→43.50/44.05→47.65).

Highlights & Insights

  • First multi-human motion editing task and benchmark: fills a gap in the field; InterEdit3D contains 5,161 high-quality triplets.
  • Frequency-domain interaction modeling: DCT decomposition with band-energy descriptors elegantly captures the rhythmic and synchronization properties of interactions.
  • InfoNCE alignment for Plan Tokens: editing intent is acquired automatically via contrastive learning without explicit annotation of which joints to modify.

Limitations & Future Work

  • Only dual-person interactions are supported; extending to three or more persons requires redesigning the interleaving strategy.
  • The model depends on the action-type coverage of the InterHuman dataset (daily activities, martial arts, dance); broader scenarios require additional data.
  • The motion representation is based on joint coordinates, lacking appearance and body-shape information.
  • MotionFix: pioneered single-person motion editing; this work extends it to multi-human settings.
  • TIMotion: the strongest multi-human generation baseline; InterEdit reuses its symmetric interleaved token design and LPA module.
  • TMR: a contrastively trained motion encoder serving as the frozen teacher for Plan Token alignment.

Rating

  • Novelty: ★★★★☆ — Task definition and frequency-domain interaction alignment are novel contributions.
  • Technical Depth: ★★★★☆ — The dual-axis Plan/Frequency Token design is complete, with a rich set of loss terms.
  • Experimental Thoroughness: ★★★★☆ — Comprehensive quantitative, qualitative, and ablation results, though all baselines are adapted rather than native methods.
  • Practicality: ★★★☆☆ — Research-driven work; dataset and code are forthcoming.

Rating

  • Novelty: Pending
  • Experimental Thoroughness: Pending
  • Writing Quality: Pending
  • Value: Pending