Seedance · 2024 – 2025
Seedance is a generative video model. My work spanned post-training and distillation: taking a capable base model, shaping it so its output is not only plausible but controllable, and then making it fast enough to actually ship.
Video is unforgiving. A single frame that drifts breaks the whole clip, and the failure is obvious to anyone watching. Most of the effort went into reward design that could see motion, not just still quality.
Sampling a video model is expensive — every extra step multiplies the cost of a clip. We distilled the model along adaptive flow trajectories and energy-matching objectives, so it could reach the same motion in far fewer steps.
The speed-ups landed in the 3.0 and 4.0 releases. The hard part was keeping the motion intact: a faster sampler that loses temporal coherence is not actually faster.
Fewer steps, same motion.
Off-the-shelf metrics reward sharp frames and punish change, which is close to the opposite of what makes video feel alive. We built an evaluation stack around two questions: does the subject persist, and does the motion read as intended?
Temporal consistency became a first-class training signal rather than an afterthought. Clips that held together scored higher even when individual frames were less pristine.
Consistency over polish.The techniques we settled on were deliberately general: preference optimization over trajectories, reward models that attend across time, and a curriculum that starts with short, simple motion.
What I took away is that post-training for generative media is closer to product design than to benchmark chasing — you are deciding what good looks like, then teaching the model to agree.
Three papers came out of the Seedance work, on diffusion acceleration and distillation. Titles link to the PDFs.