DOLLAR: Few-Step Video Generation via Distillation and Latent Reward Optimization
Ablation Study
Prompt Length (VBench)
VBench incorporates the prompt optimization technique introduced in CogVideoX,
which utilizes GPT-4o to extend the short prompts into more descriptive
"long prompts" while preserving their original meanings.
We refer to these as "long prompts", distinguishing them from the original "short prompts".
The visualization of video samples generated with short and long promtps are compared below.
With longer prompts, the text-video alignment, understanding of object relationships, and depiction of motion are generally more robust and accurate.
Short Prompt
a laptop, frozen in time
A person is baby waking up
broccoli and a backpack
In a still frame, a pear
Long Prompt
A sleek, modern laptop, its screen displaying a vibrant, paused scene, sits on a minimalist wooden desk. The room is bathed in soft, natural light filtering through sheer curtains, casting gentle shadows. The laptop's keyboard is mid-illumination, with a faint glow emanating from the keys, suggesting a moment frozen in time. Dust particles are suspended in the air, caught in the light, adding to the stillness. A steaming cup of coffee beside the laptop remains untouched, with wisps of steam frozen in mid-air. The scene captures a serene, almost magical pause in an otherwise bustling workspace.
A serene nursery bathed in soft morning light reveals a cozy crib with pastel-colored bedding. A baby, dressed in a cute onesie adorned with tiny stars, stirs gently. The camera captures the baby's delicate eyelashes fluttering open, revealing curious, sleepy eyes. The baby stretches tiny arms and legs, yawning adorably. A mobile with soft, plush animals gently spins above, casting playful shadows. The room is filled with the soft hum of a lullaby, creating a peaceful atmosphere as the baby slowly awakens, ready to greet the new day with innocent wonder.
A vibrant green broccoli floret sits atop a rustic wooden table, its fresh, crisp texture highlighted by the natural light streaming in from a nearby window. Beside it, a well-worn, navy blue backpack with leather straps and multiple pockets rests casually, suggesting a journey or adventure. The scene shifts to a close-up of the broccoli, emphasizing its intricate details and healthy appeal. Then, the camera pans to the backpack, showcasing its sturdy build and practical design. Finally, the two items are framed together, symbolizing a blend of nourishment and exploration, set against a backdrop of a cozy, sunlit room.
A single, perfectly ripe pear rests on a rustic wooden table, its golden-green skin glistening under soft, natural light. The pear's surface is dotted with tiny, delicate freckles, and its curved stem casts a gentle shadow. The background is a blurred, warm-toned kitchen scene, with hints of vintage decor and a window letting in a soft, diffused glow. The stillness of the frame captures the pear's natural beauty and simplicity, evoking a sense of calm and timelessness.