扩散强制Diffusion Forcing
Diffusion Forcing
让视频或动作序列在生成时既能一次出全片,也能逐段续写,而且不漂移。Let video or action sequences be generated all at once or extended segment by segment without drift.
训练时给每个 token 随机不同的噪声水平,模型同时学会「去噪」和「按已生成前缀续写」。
推理时可以逐帧或逐段生成,也可以并行生成;它和「一次性对整段序列做扩散」不是一回事。
它不是产品名,而是视频世界模型与实时生成常用的一类训练目标,解决长序列的一致性与延迟。
代价是训练调度更复杂:噪声水平怎么采样、前缀怎么给,都会影响长程稳定性。
During training each token gets a randomly different noise level, so the model learns both denoising and continuation from an already-generated prefix.
At inference it can generate frame by frame or segment by segment, and can also generate in parallel; it is not the same as diffusing the whole sequence in one pass.
It is not a product name but a training objective common in video world models and real-time generation, addressing long-sequence consistency and latency.
The cost is a more complex training schedule: how noise levels are sampled and how prefixes are supplied both affect long-range stability.
视频生成从「一段 5 秒」走向「30 秒、多轮延长」,这类时序一致性机制是背后的关键技术之一。Video generation is moving from five-second clips to 30-second, multi-round extensions; this kind of temporal-consistency mechanism is one of the key enablers.
回到名词地图看第4段的其他 59 个词 →Back to the map: the other 59 terms in stage 4 →