Controllable Exaggeration for Generative Motion Models via Training-Time Adaptation and Inference-Time Guidance

arXiv:2610.12316 · cs.CV · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Controllable Exaggeration for Generative Motion Models via Training-Time Adaptation and Inference-Time Guidance".

Tom: The gist The study introduces a framework for incorporating animation principles into modern motion generative models through training-time and inference-time approaches.

Jane: First, who's behind it and why it matters.

Paper summary: Tom: So we’re done with this deep dive into "Controllable Exaggeration for Generative Motion Models via Training-Time Adaptation and Inference-Time Guidance." The paper lays out a framework that handles exaggeration both during training through supervised fine-tuning on a dataset, and then at inference time using dynamic movement primitives as a guidance signal to guide the diffusion models toward exaggerated motion.

Jane: What this means for the broader field is that we are moving closer to generating character motions that aren't just technically correct in physics but also have that layer of artistic exaggeration we expect from high-quality animation. It takes those established principles, like the twelve Principles of Animation, and integrates them into the modern generative pipeline <ref:2610.12316#pg2,the 12 Principles of Animation>.

Lu: The authors are demonstrating how to map motion representation to the native space of different pre-trained motion models, which is a key step in making this work compatible across various existing systems.

Meng: From an engineering viewpoint, it shows that you can introduce a learned attribute—exaggeration—that is controllable via text prompts and then apply a mathematically defined operator for real-time adjustment during the sampling process without requiring further extensive training on the base model itself.

Lalam: The framework suggests that if we can control these motion attributes so precisely, it opens up possibilities for creating highly nuanced character expressions in media, improving how AI can generate narratives that convey deeper emotional states.

Tom: That’s the essence of it. It’s about making sure the character doesn't just move; it moves with intention and personality that an animator would design. We saw they successfully generated more exaggerated motions while keeping the underlying physical intent consistent, which is a big win for controlled generation.

Jane: And for listeners who might not be in AI research, this paper shows how we can teach machines to understand the *why* behind an animation style—the exaggeration—and apply it precisely when they are creating something new. It connects the technical math directly to the artistic goal of making characters feel alive and engaging.

Lu: The future work implied here would probably involve exploring how these DMPs interact with more complex, high-level animation concepts beyond simple scaling, maybe incorporating temporal dynamics in a more intricate way.

Meng: I think practical impact will come when we see this used for generating specific character performances that need to hit a precise level of intensity, like a sudden burst of energy or a very subtle shift in mood. That level of control is what separates good generation from production ready assets.

Tom: So, the final word on "Controllable Exaggeration for Generative Motion Models via Training-Time Adaptation and Inference-Time Guidance" is that it’s a solid blueprint for injecting established animation rules into AI, giving us motions that are both physically grounded and deeply expressive. That's our time today.

Conclusion: Tom: So we're wrapping up this look at "Controllable Exaggeration for Generative Motion Models via Training-Time Adaptation and Inference-Time Guidance." This paper really tackles how to get those professional animation principles into the stuff AI makes now.

Jane: It’s about taking that idea of exaggeration—the way animators push a motion past normal—and putting it into the machine's brain at two different times: during training, and then while it's actually making the video.

Lu: The authors show they can do this by fine-tuning the models on a dataset they made up, which includes actions at different levels of exaggeration, from zero to four.

Meng: That training stage is key because you’re teaching the model what an exaggerated version *looks* like in terms of motion data. It's giving it a head start.

Lalam: And then at inference time, they use this math—something called dynamic movement primitives—as a guide to steer the AI sampling process toward that desired exaggerated outcome without needing more training.

Tom: So, what does this actually mean for us when we’re looking at character generation? It means we can ask for a specific level of expressive energy and get it, while still keeping the original motion intent intact.

Jane: Exactly. It bridges the gap between just making something move and making it move in a way that feels intentionally animated. It lets us control the "feel" of the motion itself.

Lu: The paper proves that you can have controllable exaggeration, preserve neutral reference motion, and keep physical plausibility all at once when you use this two-stage approach.

Meng: From an engineering standpoint, it’s a clever way to inject a style attribute into a diffusion model without having to completely retrain the whole thing from scratch. It’s efficient for getting specific stylistic outputs.

Lalam: For the cultural impact, this means character design and performance generation can become much more nuanced. We can generate characters that convey complex emotions through motion in ways that feel authentic to animation history.

Tom: It's about moving beyond just physically plausible movement to something genuinely expressive. This opens up new ways for AI tools to create performances with real personality.

Jane: And while the authors show these methods work well, they also point out what the paper doesn't cover yet, which is how these concepts might scale up to even more complex animation tasks.

Lu: They flag that their current framework focuses on a few key actions and specific exaggeration levels, so extending it to every conceivable motion primitive is the next big challenge.

Meng: So, while this gets us a solid control mechanism now, the next step for engineers will be figuring out how to make that guidance signal more adaptable across different types of animation styles.

Tom: That’s where we head next. We’ll talk about how these ideas connect to the broader world of generative AI and what it means for future creative tools.

Amirhossein Zamani, Arianna Rampini, Bruno Roy

Autodesk Research · Mila – Quebec AI Institute · Concordia University

cs.CV

Submitted: 2026-10-08

Updated: 2026-10-08

Project page: https://ahhhz975.github.io/ControllableMotionExaggeration

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 87/100

The gist: The gist The study introduces a framework for incorporating animation principles into modern motion generative models through training-time and inference-time approaches.

Key concepts

Exaggeration Principle of Animation
This principle involves intentionally amplifying certain aspects of character motion to make them more expressive and engaging for the audience. The paper focuses on how to mathematically define and control this amplification so that the resulting motion feels intentional and stylized rather than just physically altered.
Training-Time Exaggeration
This stage involves fine-tuning existing text-to-motion models using a custom dataset called ExMoCap. This dataset pairs neutral actions with five levels of exaggeration, allowing the model to learn to generate motions that are already exaggerated when prompted with specific intensity descriptions.
Dynamic Movement Primitives (DMPs)
DMPs are mathematical tools used to define complex, goal-directed motion. In this framework, DMPs are used to create a differential equation that models how a neutral motion should evolve, which then allows the system to scale and exaggerate the resulting movement by modifying its forcing term.
Inference-Time Guidance
This technique applies the learned mathematical formulation of exaggeration during the actual generation process. Instead of retraining the model, this guidance signal steers existing diffusion or flow-matching models to produce exaggerated results without requiring any additional training data.

Terminology

Summary

The gist The study introduces a framework for incorporating animation principles into modern motion generative models through training-time and inference-time approaches.

Problem Statement and Preliminaries

Character animation relies on established principles like the 12 Principles of Animation to create expressive character motions, but recent motion generative foundation models often neglect these principles in their generation pipeline. To address this gap, the authors target the Exaggeration principle of animation to generate motions that are exaggerated and expressive. Incorporating the Exaggeration principle into a motion generative pipeline requires understanding how animators interpret exaggeration and how they apply it in practice when designing expressive motions for characters. As a result of these discussions, the authors identify three properties that an exaggerated motion generation method should satisfy: (i) controllable exaggeration; (ii) motion-intent preservation; and (iii) physical plausibility.

Training-Time Exaggeration

The first stage of the framework introduces exaggeration during training through supervised fine-tuning of pre-trained text-to-motion models on a curated dataset. The authors construct an ExMoCap dataset containing 26 actions with five exaggeration levels per action, where level 0 corresponds to the neutral version and level 4 corresponds to its most exaggerated version. They perform supervised fine-tuning using this dataset where each discrete level is associated with an intensity description appended to the underlying action caption The authors use techniques like data augmentation including left-right mirroring and position jittering to expand the ExMoCap dataset. To mitigate overfitting during fine-tuning on smaller datasets, they adopt three techniques: data augmentation, low-rank adaptation (LoRA), and prior preservation.

Inference-Time Exaggeration

The second stage operates at inference time where they introduce exaggeration without modifying the parameters of the pre-trained model. This involves two steps: formulating the Exaggeration principle mathematically using DMPs, and using this formulation as a guidance signal for diffusion or flow-matching text-to-motion models.

Mathematical Formulation of Exaggeration

The authors formulate exaggeration as a geometric transformation of a neutral motion, defining an exaggeration operator QK = EK(Qneut), where K is a scalar controlling the exaggeration. They leverage dynamic movement primitives (DMPs) to define this operator The DMP formulation involves a differential equation that combines a goal-directed spring-damper term and a learned forcing term. Exaggerating motion is achieved by scaling this forcing term by a factor K: fK(x) = Kf(x). In diffusion models, they modify the sampling process by moving the prediction Qˆ(s) toward the DMP-exaggerated target QK before each update

Evaluation Metrics

They propose a dedicated evaluation suite for quantifying exaggeration in generated motions, consisting of classifier-based and classifier-free metrics. Classifier-based metrics include ELRA@k and ρclf. Classifier-free metrics include root-relative range of motion (ROM) as a physical-intensity score.

Conclusion

The study introduces a framework for incorporating the Exaggeration principle of animation into modern motion generative models through training-time and inference-time approaches. Qualitative and quantitative evaluations demonstrate that their methods generate more expressive motions while preserving the underlying action.

Improvements for AI systems

  1. Bold motion generation models will now produce motions that are more expressive and exaggerated while preserving intent, by leveraging a two-stage framework: training-time fine-tuning on an ExMoCap dataset for learning exaggeration as a text-conditioned attribute, followed by inference-time guidance using a DMP formulation.

  2. The system will be able to generate controllable motion variations where the user can adjust intensity via text prompts, as demonstrated by Figure 11 and Table 5, allowing the generation of motions that are progressively larger and more expressive poses while preserving the underlying neutral action.

  3. Inference-time guidance will allow existing diffusion and flow-matching models to be steered toward exaggerated motion without additional training, using a formulation where exaggeration becomes a learned text attribute guided by dynamic movement primitives (DMPs) to produce motion without modifying the parameters of the pre-trained model.

  4. The system will possess robust action preservation capabilities quantified by metrics like R-precision@k and MMDist, ensuring that exaggerated motions still adhere to the original action, as evidenced by Table 1 showing that SFT+TF achieves a high score in these areas compared to baselines.

  5. The AI can perform a comprehensive evaluation of motion quality using a dedicated suite of metrics, including classifier-based metrics like ELRA@k and classifier-free metrics like root-relative range of motion (ROM) to quantify both the degree of exaggeration and the physical plausibility of the output.

  6. The system can be fine-tuned across different generative backbones (MDM, HY-Motion, KIMODO) while maintaining a consistent trade-off between exaggeration control and prompt adherence by selecting balanced operating points for LoRA rank and prior-preservation weight, as detailed in Table 6 and Table 8 ablation studies.

  7. The system will be able to generate motions that are physically plausible despite high exaggeration levels because the inference process uses a pre-trained model's learned physical motion prior to iteratively refine the DMP-exaggerated motion, preventing artifacts like invalid body poses.

Abstract

Recent motion generative models have demonstrated strong capabilities in synthesizing physically plausible character motion, but often overlook established animation principles used by professional animators to ground and design their animation work. Understanding and incorporating these principles into motion generative pipelines is essential for producing motions that serve not only physically grounded applications but also the needs of the character animation community. This enables the creation of characters that not only move in physically plausible ways but also feel alive, expressive, and engaging. To close this gap, we focus on the Exaggeration principle of animation and investigate how it can be incorporated into modern motion generative pipelines to produce more expressive character motions. To this end, we introduce a framework that operates at two stages of existing motion generative pipelines. The first stage introduces exaggeration during training, where we perform supervised fine-tuning of pre-trained text-to-motion models on our curated exaggeration dataset. The second stage operates at inference time, where we: (i) introduce a mathematical formulation of exaggeration based on dynamic movement primitives (DMPs); and (ii) leverage this formulation as an exaggeration guidance signal to guide existing diffusion and flow-matching text-to-motion generation models toward exaggerated motion without additional training. Through qualitative and quantitative evaluations against three strong motion generation models, we show that our methods generate more exaggerated and expressive motions while preserving neutral reference motion intent and physical plausibility.

Sources

Related papers