Researchers extend APEX to few-step streaming video generation without a fake-score network
A preprint on arXiv presents APEX, a method for few-step streaming video generation that drops the separate fake-score network normally used in distribution matching distillation. The generator estimates the velocity field of its own output by passing a shifted version of the text condition through the same parameters, so the original condition drives generation while the shifted one drives the estimate. S2-Avatar's frozen reference teacher, Self-Replay Forcing and perceptual regularization are kept.
The authors evaluated APEX on 180-second videos with a 70-case human study and a 320-case StreamAV-Bench comparison. Against S2-Avatar it scored higher on instruction fulfillment (3.856 versus 3.831) and visual aesthetics (0.641898 versus 0.632606); raters preferred it by 30.0% net on instruction following, with consistency and video quality at +4.3% each. Under matched training settings it used 4.4% less peak allocated GPU memory per GPU and 9.2% shorter mean iteration time than DMD.
Why it matters
Few-step sampling is how streaming video generation gets cheap, but the extra fake-score network adds training memory and compute and has to track a moving generator. This approach removes that network while keeping memory and iteration time below DMD, though human raters still favored S2-Avatar on motion quality.
Sources
- PrimaryHugging Face Papers
Written by AIToolMaps from the source above and checked against it sentence by sentence.