Few-Step Streaming Video Generation via Condition Shifting
APEX removes a separate training network, cutting peak GPU memory by 4.4% and mean iteration time by 9.2% against DMD under matched settings.
The method uses shifted text conditioning to make the generator itself estimate the generated distribution, while retaining a frozen reference teacher. Tests covered 180-second videos, a 70-case human study and a 320-case benchmark comparison. Human raters preferred APEX over S2-Avatar most strongly for instruction following, with a +30.0% net preference; consistency and video quality each gained +4.3%. Motion quality favored S2-Avatar, audiovisual synchronization was predominantly tied, and usability was unchanged.
HF Daily Papers' note
The method uses shifted text conditioning to make the generator itself estimate the generated distribution, while retaining a frozen reference teacher. Tests covered 180-second videos, a 70-case human study and a 320-case benchmark comparison. Human raters preferred APEX over S2-Avatar most strongly for instruction following, with a +30.0% net preference; consistency and video quality each gained +4.3%. Motion quality favored S2-Avatar, audiovisual synchronization was predominantly tied, and usability was unchanged.
HF Daily Papers' note
score 5