Abstract
Despite recent advancements 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, which 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. Combined with a piecewise guidance schedule, MotionCFG consistently improves motion dynamics across state-of-the-art T2V frameworks with negligible computational overhead.
Methodology
MotionCFG injects Gaussian noise exclusively into motion-related concept embeddings (e.g., action verbs) to create localized negative anchors. By repelling the generation trajectory away from this motion-ambiguous manifold during the early denoising steps, our method resolves the inherent static bias of T2V models without distorting spatial structures.
1. Enhancing Motion Dynamics in Text-to-Video Generation
Compared to negative-prompted baselines and state-of-the-art diversity samplers (CADS, IG), MotionCFG produces physically plausible and rigorous dynamics faithful to the prompt, completely avoiding static collapse.
"A blue car drives past a white picket fence on a sunny day"
"Man in a red suit driving a black sports car"
"Porcelain rabbit hopping by a golden cactus"
"A pink teddy bear sits on a black oven in the kitchen"
"A lady wearing a pink dress holds a green umbrella, twirling it as she dances in the rain"
"A girl in a black dress watering a green rose in the garden"
"An orange tractor pulling a cart of purple carrots"
"A dog runs through a field while a cat climbs a tree"
"A monkey plays with fruit, a deer grazes nearby"
"A duck swims in the pond and a frog hops onto a lily pad"
"A man jogs and a car drives on the road"
"A woman rides a bicycle and a drone hovers above capturing the scene"
"A dog chases after a balloon drifting in the wind"
"A cat skates on a skateboard, chasing a rolling ball"
"A horse races alongside a moving train"
"A boy plays basketball and a girl jumps rope"
"A rabbit ice skates on a frozen lake while a chipmunk builds a snowman"
"A golden retriever running fast across a green field"
"A wild horse galloping through a dusty canyon"
"A futuristic car explodes into pieces on the highway"
"A robot transforming into a car"
"A volcano erupting with lava and ash"
"Clouds moving rapidly across a blue sky"
"An old building collapsing into dust"
2. Improving Object Counting in Text-to-Video Generation
Beyond motion, the noise-induced contrastive mechanism of MotionCFG can be generalized to steer non-linear semantic attributes such as precise object counting, maintaining structural integrity throughout the video sequence.
"Three children run along the tracks."
"Four dogs running through a field."
"Four bears lumber through the woods."
"Four chairs rock on a porch, creaking with the breeze."
"Five cats walking on the floor."
"Five rabbits hop through a meadow."
"Five pandas munch bamboo in the garden."
"Five cows amble through the pasture."
"Three apples hang heavy on the tree branch, and three birds perch nearby."
