Ignoring denoising stages when quantizing video models ruins textures and sharpness.

Computer programs that generate video from text take huge amounts of memory and computing power. To make them smaller and faster, researchers use quantization, which simplifies the model's numbers. But existing methods blur fine textures, visual details, and sharpness. The researchers found that standard compression treats every step of making a video the same way. In reality, early steps plan basic shapes and movement, while later steps add sharp details. To fix this, the researchers created DSAQuant. During training, their method learns structure from a full-sized teacher model in early steps, then focuses on building fine details in later steps. When generating videos, it turns off guidance settings at the very end so tiny errors are not blown up into weird visual flaws. In tests, DSAQuant generated sharper videos and clearly beat the top existing method, keeping strong alignment between prompts and videos.

Why it matters

Video generation models are too slow for many real-world uses because they need lots of memory. This paper makes them faster and lighter without making the videos blurry or lose detail, which helps anyone using AI video tools.

Questions answered