few-step schedule

A few-step schedule is a work sequence that minimizes the number of 'passes' an AI image-making model needs before it delivers a finished piece. Video and image generation models typically start from a screen of static-like noise and repeat the process of removing noise dozens of times to complete an image—similar to a painter building up an artwork by layering paint over a sketch again and again.

A few-step schedule tells this painter to make do with only four passes, and to settle everything—from the rough sketch to the final touches—within those four. Painting four random strokes in any order would leave the picture messy, but if the painter follows a pre-calculated sequence of exactly where and how boldly to apply the brush at each stage, a convincing finished piece can emerge even with so few passes. This pre-set sequence is the schedule.

As a result, fewer iterations mean less computation time, making it possible for the time it takes to generate a video to become shorter than the video's own runtime.

The article explains that they implemented a few-step schedule in two different ways, 'cutting it down to just 4 DiT computations across 5 noise-intensity points (sigma points).' What's easy to misunderstand here is that the schedule itself isn't the magic that produces speed—it's simply the plan that specifies where and how many times to apply a 'pass.' The actual speedup comes from applying this schedule on top of an implementation method such as Turbo LoRA or FastH3.