Person-Interaction Noise Optimization (PINO)

• Person-Interaction Noise Optimization (PINO) is a training-free framework that decomposes multi-person motion synthesis into semantically meaningful pairwise interactions using pretrained diffusion models.
• It employs a modular factorization strategy by repeatedly applying a frozen two-person model and optimizing latent noise with physics-based penalties to ensure physical coherence and precise user control.
• Empirical results report dramatic reductions in overlap and penetration errors, demonstrating enhanced motion realism and controllability in both two-person and larger group settings.

Searching arXiv for the specified papers and closely related work. arxiv_search: {"query":"id:(Ota et al., 25 Jul 2025) OR id:(Zhang et al., 1 Feb 2025) OR InterGen two-person diffusion motion generation", "max_results": 10, "sort_by": "submittedDate", "sort_order": "descending"} Person-Interaction Noise Optimization (PINO) is a training-free framework for generating realistic and customizable interactions among groups of arbitrary size by decomposing a multi-person motion synthesis problem into semantically relevant two-person interactions and then optimizing latent noises under pretrained diffusion priors and physics-based penalties (Ota et al., 25 Jul 2025). In the formulation reported for long-duration group motion generation, PINO uses a frozen two-person interaction model, composes larger scenes incrementally, and exposes differentiable controls over orientation, speed, and spatial relationships without additional training. Its reported purpose is to improve physical coherence and user controllability relative to conditional group-generation procedures that rely on a single shared prompt.

1. Problem setting and core representation

PINO treats group interaction synthesis as a composition problem over pairwise interaction priors. Let $M$ denote the final number of people, $x^p\in\mathbb R^{N\times D}$ the $N$-frame motion of person $p$, $z^p\in\mathbb R^{N\times D}$ the corresponding initial noise, and $c_{i,j}$ the text prompt for the interaction between persons $i$ and $j$. The pretrained two-person diffusion model is written as a conditional density

$$p_\theta\bigl(x^i,x^j \mid z^i,z^j,c_{i,j}\bigr) = \mathcal{N}\Bigl((x^i_0,x^j_0);\mu_\theta(z^i,z^j,c_{i,j}),\Sigma\Bigr),$$

or equivalently in denoiser form

$$[\,x^i_0,x^j_0\,] = G_\theta\bigl(z^i,z^j,c_{i,j}\bigr).$$

Within this representation, a multi-person scene is not modeled by a monolithic $x^p\in\mathbb R^{N\times D}$0-person generator. Instead, PINO uses repeated application of a pretrained two-person model $x^p\in\mathbb R^{N\times D}$1. This suggests a modular factorization of group interaction semantics: pairwise prompts provide the semantic scaffolding, while the optimization over the full latent set $x^p\in\mathbb R^{N\times D}$2 reconciles these local interaction constraints into a coherent group motion. The paper positions this design against existing conditional diffusion approaches that incrementally generate motions by conditioning on previously generated characters but rely on single shared prompts, which limits nuanced control (Ota et al., 25 Jul 2025).

2. Pairwise decomposition and latent-noise optimization

The central objective jointly optimizes the latent noises for all persons. PINO minimizes a sum of reconstruction losses over prompted pairwise interactions together with a physics-based penalty:

$x^p\in\mathbb R^{N\times D}$3

Here $x^p\in\mathbb R^{N\times D}$4 is the sum of DDIM reconstruction errors for each prompted pair, $x^p\in\mathbb R^{N\times D}$5 aggregates physical plausibility terms, and $x^p\in\mathbb R^{N\times D}$6 balances physical versus semantic fidelity.

The composition algorithm is greedy and incremental. PINO takes as input an ordered list of participants $x^p\in\mathbb R^{N\times D}$7, and for each new person $x^p\in\mathbb R^{N\times D}$8, a reference index $x^p\in\mathbb R^{N\times D}$9 and a prompt $N$0. It first generates and optimizes the base pair $N$1. Then, for each subsequent participant $N$2, it samples $N$3, fixes the optimized reference motion $N$4, and optimizes the new latent with respect to pairwise fidelity and physics constraints before appending the resulting motion to the group. In the reported pseudocode, the incremental objective for person $N$5 is

$N$6

A notable feature is that the diffusion backbone remains frozen. Rather than retraining the generator for larger groups, PINO performs gradient-based optimization on initial latents. Each optimization step backpropagates through the DDIM denoising chain $N$7, which holds the reference person’s denoised path fixed by injecting noise back into it and only denoises the target’s latent to produce the pairwise motion (Ota et al., 25 Jul 2025). A plausible implication is that the method trades parameter adaptation for latent-space search, which is consistent with its description as training-free.

3. Physics-based penalties and physical coherence

The physics term is defined as

$N$8

Its purpose is to suppress common artifacts such as overlapping and penetration between characters while preserving temporally plausible transitions.

The overlap term is a root-distance hinge loss:

$N$9

The penetration term is mesh-based:

$p$0

where $p$1 is the interpenetration volume of fitted SMPL meshes. The boundary term penalizes acceleration spikes in transition regions:

$p$2

with $p$3 denoting the first few inpainted frames.

These penalties define the mechanism by which PINO seeks physical coherence. The overlap term acts on coarse spatial separation, the penetration term on fine-grained body interpenetration, and the boundary term on motion smoothness. This layered penalty design indicates that physical plausibility is not treated as a single scalar criterion but as a conjunction of collision avoidance and temporal regularity (Ota et al., 25 Jul 2025).

4. Customization and controllable generation

PINO exposes user control through additional differentiable losses. The reported control objective is

$p$4

The available controls are desired root-trajectory waypoints, movement-region constraints, facing-direction targets, and desired inter-character distances or angles. These terms are appended to the overall objective and weighted by user-supplied $p$5 coefficients.

The paper explicitly describes these controls as operating on orientation, speed, root position, and inter-character spacing, with examples such as a hinge loss on the distance between actual and target root positions and a max-violation penalty for orientation. This control structure is central to PINO’s claim of customizability: user intent is encoded not by retraining or prompt engineering alone, but by adding task-specific differentiable regularizers to latent optimization (Ota et al., 25 Jul 2025). This suggests a separation between semantic priors, which come from the pretrained two-person model, and scene-specific constraints, which are enforced through optimization.

5. Implementation regime and reported empirical results

The reported implementation uses InterGen as the base two-person diffusion model, a 50-step DDIM sampler, and Adam optimization on latents with learning rate $p$6 for 100 steps, with early stopping at loss $p$7. The overlap threshold is $p$8 cm, the orientation threshold is $p$9 as a cosine threshold, the acceleration penalty is applied on the first 25 frames of an extension, and $z^p\in\mathbb R^{N\times D}$0, while other $z^p\in\mathbb R^{N\times D}$1 values are set per task.

The evaluation protocol separates non-semantic and semantic criteria. The non-semantic metrics are overlap rate, defined as the fraction of frames with root-distance $z^p\in\mathbb R^{N\times D}$2 cm, penetration volume in cm$z^p\in\mathbb R^{N\times D}$3, foot skating error, and maximum acceleration. The semantic metrics are FID against ground-truth motions, Diversity as sample variance, and R-Precision for text-motion retrieval.

The reported highlights are specific. In two-person settings, overlap dropped from $z^p\in\mathbb R^{N\times D}$4 to $z^p\in\mathbb R^{N\times D}$5 and penetration from $z^p\in\mathbb R^{N\times D}$6 cm$z^p\in\mathbb R^{N\times D}$7 to $z^p\in\mathbb R^{N\times D}$8 cm$z^p\in\mathbb R^{N\times D}$9 with PINO. In five-person settings, overlap remained $c_{i,j}$0 for all pairs, whereas FreeMotion and InterGen exceeded $c_{i,j}$1. For motion extension, FID improved from $c_{i,j}$2 to $c_{i,j}$3, with lower foot skating error and acceleration. A user study with 35 participants gave PINO the highest Text Fidelity and Interaction Quality and the most overall votes (Ota et al., 25 Jul 2025). The abstract further states that the resulting motions are visually realistic, physically coherent, and adaptable, and suitable for animation, gaming, and robotics applications.

6. Terminological ambiguity and relation to other uses of “PINO”

The acronym “PINO” is not unique in the 2025 arXiv literature. In recommender systems, the paper “Personalized Denoising Implicit Feedback for Robust Recommender System” uses the label “PLD/PINO” for a resampling strategy based on a user’s personal loss distribution $c_{i,j}$4, where low-loss interactions are preferentially resampled through a softmax-style distribution

$c_{i,j}$5

That method addresses noisy implicit feedback in recommendation rather than human motion generation, although both usages involve optimization over noisy signals (Zhang et al., 1 Feb 2025).

The recommender-system PINO operates by candidate-pool construction and loss-informed resampling, with a theoretical claim that expected bias toward normal interactions increases when $c_{i,j}$6, and with experiments on Gowalla, Yelp2018, MIND, and MIND-Large using MF and LightGCN under BPR and BCE losses. By contrast, Person-Interaction Noise Optimization is a motion-generation framework built around pretrained two-person diffusion models, latent optimization, and physics-based penalties (Zhang et al., 1 Feb 2025). This naming overlap can create bibliographic ambiguity; a plausible implication is that citation by full title or arXiv identifier is preferable when disambiguation matters.