Abstract
Despite recent advances in Text-to-Video (T2V) synthesis, generating high-fidelity and dynamic motion remains a significant challenge. Existing methods primarily rely on Classifier-Free Guidance (CFG), often with explicit negative prompts (e.g. "static", "blurry"), to suppress undesired artifacts. However, such explicit negations frequently introduce unintended semantic bias and distort object integrity; a phenomenon we define as Content-Motion Drift. To address this, we propose MotionCFG, a framework that enhances motion dynamics by contrasting a target concept with its noise-perturbed counterparts. Specifically, by injecting Gaussian noise into the concept embeddings, MotionCFG creates localized negative anchors that encapsulate a broad complementary space of sub-optimal motion variations. Unlike explicit negations, this approach facilitates implicit hard negative mining without shifting the global semantic identity, allowing for a focused refinement of temporal details. Combined with a piecewise guidance schedule that confines intervention to the early denoising steps, MotionCFG consistently improves motion dynamics across state-of-the-art T2V frameworks with negligible computational overhead and minimal compromise in visual quality. Additionally, we demonstrate that this noise-induced contrastive mechanism is effective not only for sharpening motion trajectories but also for steering complex, non-linear concepts such as precise object numerosity, which are typically difficult to modulate via standard text-based guidance.
Introduction: The Challenge of Static Bias
Current Text-to-Video (T2V) models, despite their impressive realism, suffer from a persistent "static bias." Because training datasets like WebVid-10M are dominated by relatively still shots or subtle camera pans, models often default to generating nearly static videos even when prompted for high-intensity action (e.g., "a cat jumping over a fence").
To counteract this, researchers typically employ Classifier-Free Guidance (CFG) with explicit negative prompts such as "static" or "blurry." However, this approach leads to a significant side effect termed Content-Motion Drift. When a model is instructed to avoid being "static," it often distorts the structural integrity or identity of the subject to create perceived movement, resulting in morphing objects or loss of detail.
Methodology: Stochastic Concept Perturbation
MotionCFG introduces a training-free, "plug-and-play" framework to enhance motion dynamics through three primary steps:
- Motion Token Identification: A lightweight LLM (e.g., Qwen2.5-7B) identifies "motion-critical" tokens in the prompt—typically action verbs like "galloping," "spinning," or "collapsing."
- Stochastic Concept Perturbation: Instead of a generic negative prompt, MotionCFG injects Gaussian noise into the embeddings of these specific motion tokens. This creates a "localized negative anchor" representing a mathematical representation of a "degraded" or "sub-optimal" version of that specific action.
- Piecewise Guidance Schedule: Diffusion models establish global layout and motion in early denoising steps and refine textures later. MotionCFG is applied only during the initial 10-20% of the process to establish the motion trajectory, after which the model reverts to standard CFG for visual polish.
By contrasting the clean motion embedding with the noisy one, the model is repelled from the ambiguous, low-motion space and forced toward a more distinct, dynamic trajectory without shifting the global semantic identity of the objects.
Key Results and Insights
The researchers evaluated MotionCFG across state-of-the-art frameworks including Wan2.1 (1.3B and 14B) and CogVideoX-5B.
- Quantitative Performance: On Wan2.1-14B, MotionCFG more than doubled the FLOW score (pixel-level motion) compared to the baseline (4.48 vs. 1.71) and outperformed diversity-enhancing methods like CADS on the DEVIL score.
- Efficiency: The method is training-free and adds only approximately 1 second of overhead per video generation on an NVIDIA B200 GPU.
- Beyond Motion: Surprisingly, the noise-induced contrast mechanism also improved object counting accuracy (e.g., generating exactly "three" objects), demonstrating its utility for other non-linear concepts difficult for standard text-based guidance.
Conclusion
MotionCFG demonstrates that implicit hard negative mining via noise perturbation is a superior alternative to explicit text-based negations for video synthesis. By focusing on motion-specific tokens and employing a piecewise schedule, it successfully boosts kinetic energy in AI-generated videos while preserving semantic identity and object integrity.
Summary updated