acceptodds
Under review as a conference paper at ICLR 2027

Single Stream, Single Step: Simplifying 3D Human–Human Interaction Generation

Abstract

Generating realistic 3D human–human interactions requires plausible individual motion and coherent coordination between two individuals. Existing approaches often use separate per-person streams or specialized interaction modules, while some additionally rely on learned motion autoencoders. These designs complicate the architecture or training pipeline, and iterative sampling adds inference cost. We introduce InterFlow, an end-to-end autoregressive flow-matching model that simplifies interaction generation using a single-stream diffusion transformer. InterFlow operates directly in motion space without a variational autoencoder, representing the motions of both individuals as a unified token sequence and modeling temporal and interpersonal dependencies through shared self-attention. Its fixed-window formulation enables blockwise generation with a computational cost per block independent of the total generated sequence length. We further introduce FastInterFlow, which combines rollout-based score distillation with forward-kinematics-based adversarial distillation to obtain a one-step generator from the multi-step teacher. FastInterFlow retains the autoregressive formulation and generates each motion block in a single network evaluation. Empirical results demonstrate the effectiveness of InterFlow’s streamlined architecture for modeling interactions. Together, the two models provide a simple framework combining unified motion modeling with efficient blockwise autoregressive generation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Human-human interaction generation, autoregressive motion generation, one step distillation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_WGW7OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents InterFlow, a single-stream autoregressive flow-matching framework for text-driven 3D human-human interaction generation. Instead of modeling the two interactants using separate branches or a learned latent motion representation, the method directly operates in raw SMPL-X motion space and represents the two persons as interleaved person-frame tokens. A factorized 2D positional encoding is used to distinguish temporal positions and person identities, while shared self-attention jointly models individual motion dynamics and interpersonal coordination. The paper further proposes FastInterFlow, which distills the multi-step InterFlow model into a one-step autoregressive generator. The distillation combines rollout-based score distillation with a forward-kinematics-based adversarial objective defined on joint positions and velocities. Experiments are conducted on Inter-X, where the proposed methods show competitive or improved performance over InterGen, InterMask, and Interact2Ar, together with substantial inference speedup.

Critical strengths

Strengths

  1. The overall framework is simple and practically appealing. The idea of treating both persons as a unified token sequence avoids complicated handcrafted interaction modules, and the proposed single-stream architecture is conceptually clean.
  2. The paper pays serious attention to efficiency. The fixed-window autoregressive formulation bounds the computation of each generation block, and FastInterFlow further reduces iterative flow sampling to a single network evaluation. This is useful for long or online interaction generation.
  3. The experimental analysis is relatively detailed. The authors study positional encoding, dual-stream versus single-stream modeling, history allocation, diffusion forcing, prediction parameterization, and distillation objectives. In particular, the architectural ablation suggests that the proposed shared-stream design is effective in the authors' implementation.
  4. The consideration of articulated hands is valuable. Compared with many earlier HHI methods focusing mainly on body motion, evaluating Full/Body/Hands separately on Inter-X provides useful evidence regarding fine-grained motion generation.
Critical weaknesses

Major weaknesses

  1. Missing comparison with recent latent-space HHI methods.

One of the main motivations of this paper is that learned autoencoder representations may introduce an information bottleneck and potentially affect fine-grained interaction details. In this regard, I think DHVAE (ICLR 2026) [1] is a particularly relevant recent work that deserves discussion. DHVAE also studies 3D human-human interaction generation, but takes a different direction by designing structured and disentangled latent representations. Therefore, comparing or at least discussing these two design choices would make the motivation of the proposed raw-motion formulation much clearer. I do not necessarily think a completely matched comparison is mandatory if there are implementation or representation differences, but I would strongly encourage the authors to clarify the relationship with DHVAE and, if possible, provide an empirical comparison.

  1. The novelty of one-step autoregressive generation needs clearer positioning.

The paper describes FastInterFlow as the first one-step generator for autoregressive interaction synthesis. However, ARMFlow (CVPR 2026) [2] studies one-step autoregressive generation in the closely related setting of online 3D human reaction generation. I understand that reaction generation, where one actor serves as the conditioning signal, is not exactly the same as jointly generating two interacting humans. Nevertheless, both works investigate one-step generation, autoregressive rollout, and error accumulation in two-person motion settings. A discussion distinguishing FastInterFlow from ARMFlow would make the novelty claim much clearer. If technically feasible, a comparison under a common benchmark would also be informative.

  1. The modeling of physical contact is not sufficiently clear.

The paper emphasizes the importance of physical contact and states that FK adversarial distillation helps enforce physical plausibility and contact coordination. However, from my understanding, the discriminator mainly receives world-space joint positions and joint velocities, without an explicit contact representation, contact label, inter-person distance loss, or penetration/collision objective. Therefore, I am not fully clear whether contact is expected to emerge implicitly from the real/fake kinematic distribution, or whether there is another contact-related mechanism that I may have missed. It would also strengthen the paper if the authors could report interaction-specific physical metrics, such as contact accuracy, missed-contact rate, penetration, or collision statistics.

  1. Evaluation on only Inter-X is somewhat limited

All major experiments are currently conducted on Inter-X. While Inter-X is suitable for evaluating detailed body and hand motions, InterHuman is also a widely adopted benchmark for text-driven human-human interaction generation. Since the proposed single-stream formulation does not appear to fundamentally depend on the Inter-X representation, I am curious why InterHuman is not included. Evaluation on InterHuman would enable comparison with a broader set of existing HHI methods and would also help demonstrate that the benefit of the proposed architecture generalizes across datasets. I would encourage the authors to provide InterHuman results if possible, or at least clarify the reason for focusing exclusively on Inter-X.

  1. Some architectural and long-horizon claims would benefit from more targeted experiments

The single-stream formulation is simple and elegant, and the dual-stream ablation provides useful evidence. However, the conclusion that specialized interaction modules are unnecessary seems broader than what can be established from one internally implemented dual-stream baseline. In addition, long-horizon autoregressive generation is an important motivation of the paper, but most reported results are aggregate generation metrics. It would be helpful to report performance as rollout length increases, or provide measurements of block-boundary continuity, accumulated motion errors, or accumulated interaction/contact errors. Such experiments would better support the claimed advantages of the autoregressive design.

[1] Geng et al. "Disentangled Hierarchical VAE for 3D Human-Human Interaction Generation." International Conference on Learning Representations (ICLR), 2026. [2] Geng et al. "ARMFlow: AutoRegressive MeanFlow for Online 3D Human Reaction Generation." IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.

Questions for authors

The authors are encouraged to:

  1. Could the authors clarify the differences and comparisons with DHVAE [1]?
  2. As far as I know, ARMFlow [2] is the first single-step generation in HHI interaction; could the authors clarify this?
  3. Could the authors clarify the contact modeling in InterFlow? Results with Physical metrics may help demonstrate this.
  4. Since the authors claim that their approach benefits long-horizon generation, could the authors conduct experiments on the InterHuman dataset, which has longer sequence lengths? (maximum 300 frames)

[1] Geng et al. "Disentangled Hierarchical VAE for 3D Human-Human Interaction Generation." International Conference on Learning Representations (ICLR), 2026. [2] Geng et al. "ARMFlow: AutoRegressive MeanFlow for Online 3D Human Reaction Generation." IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.

Reviewer_kV2eOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes InterFlow, an autoregressive flow-matching framework for text-conditioned 3D human–human interaction generation. It represents both participants as a unified sequence of person–frame tokens and models their coordination through shared self-attention, without a motion autoencoder. A fixed-window formulation supports blockwise generation with bounded computation per block. The paper further introduces FastInterFlow, which combines rollout-based score distillation and forward-kinematics adversarial supervision to generate each block in one network evaluation. Experiments on Inter-X demonstrate competitive body and hand generation metrics, substantial inference acceleration, and the effectiveness of several architectural and training choices.

Critical strengths
  1. The unified person–frame representation models both participants using a shared transformer, removing separate person-specific branches and autoencoder training. The comparisons and architectural ablations in Tables 1–2 support the effectiveness of this simplified design, with strong results across full-body, body-only, and hand evaluations.
  2. FastInterFlow reduces the reported test-set generation time from 97 seconds to 15 seconds relative to its teacher while maintaining competitive generation metrics. The distillation ablations demonstrate the importance of student training and clarify the contribution of FK adversarial supervision. This provides a practically useful quality–efficiency tradeoff for autoregressive interaction synthesis.
Critical weaknesses
  1. The motivation for single-stream modeling is insufficiently developed relative to prior work. The paper does not clearly explain why this design should improve interaction modeling, and its implications for independently controlling each participant remain unclear.
  2. Although the overall results are strong, Table 1 shows worse hand-generation performance when ground-truth first-frame conditioning is added. This behavior requires clearer explanation, including whether the compared configurations use consistent settings and evaluation protocols.
Questions for authors
  1. The advantages of single-stream modeling over previous designs need clearer justification, particularly regarding its ability to support independent control of each participant.
  2. The worse hand results with first-frame ground-truth conditioning in Table 1 require explanation. The comparability of the configurations and variability across existing evaluation runs should also be clarified.
  3. Evaluation is limited to Inter-X, providing limited evidence of generalization beyond this dataset. The scope of the conclusions should reflect this limitation.
  4. The production process and intended role of Figures 1–3 should be clarified, including any use of generative tools. Conceptual illustrations should be distinguished from actual model outputs.
  5. The conclusion appears outside the main text. Its placement should be checked against the venue’s formatting and page-limit requirements.
Reviewer_9t8YOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses text-conditioned two-person motion generation, where coordinating body and hand movements can require complex architectures and costly sampling, by using a shared motion transformer and distilling it into a one-step autoregressive generator.

Critical strengths
  1. The unified model handles both participants’ body and hand motions without dedicated trajectory, body and hand branches or a learned motion autoencoder, simplifying the organization of interaction generation.
  2. One-step block generation reduces the repeated sampling work needed to produce each subsequent motion segment, offering a practical efficiency advantage for sequential animation.
Critical weaknesses
  1. The paper attributes motion fragmentation to separate branches and loss of fine contact details to motion autoencoders, but the reported comparisons mainly establish the performance of the overall system rather than isolating either cause. (Section 1; Section 4.3.)

  2. The FK adversarial stage introduces additional training, while it is unclear whether the score-only comparison uses the same training duration and rollout schedule, so the additional gain cannot be attributed specifically to FK supervision. (Sections 3.3 and 4.3.)

  3. Improved feature-distribution metrics do not establish better physical contact or subtle interpersonal coordination, so the quality claims for the one-step model need to distinguish distributional fidelity from the quality of the interaction itself. (Sections 4.2–4.3.)

Questions for authors

No

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.