Question about Self-Forcing-Plus

[yjhong89]

Hi!

I am looking into Self-Forcing-Plus repository implementing self-forcing with dmd.

• https://github.com/GoatWu/Self-Forcing-Plus/tree/wan22

@GoatWu

• There are 2 training pipelines supported in SelfForcinModel, and wonder how SelfForcingTrainingPipeline and BidirectionalTrainingPipeline work differently ??
  • https://github.com/GoatWu/Self-Forcing-Plus/blob/f947063ef694ade8f212c1806aed2686bdfa721a/model/base.py#L229
  • Why generator_type set differently??
    • https://github.com/GoatWu/Self-Forcing-Plus/blob/f947063ef694ade8f212c1806aed2686bdfa721a/configs/self_forcing_14b_i2v_dmd.yaml#L8
    • https://github.com/GoatWu/Self-Forcing-Plus/blob/f947063ef694ade8f212c1806aed2686bdfa721a/configs/self_forcing_dmd.yaml#L7
• Why boundary of high and low noise set up to 0.5? Original wan uses this value as 0.9
  • https://github.com/GoatWu/Self-Forcing-Plus/blob/f947063ef694ade8f212c1806aed2686bdfa721a/configs/wan22_high_i2v.yaml#L16
• Why recent trends use self-forcing with dmd, not using only dmd without self-forcing or any causality?
  • Though read from docs, since DMD was originally supposed to handle image generation, so adopt self-forcing ? (training stability ?)
  • Still don't get it why self-forcing (or causvid) is naturally combined with DMD (or other distillation method). Any advice?

• If I train self-forcing (w.DMD) algorithm to bidirectional algorithm (e.g. Wan2.1/Wan2.2), then can distilled model act like an auto-regressive video generator??
• What is difference between step-distllation and autoregressive distillation in docs?
  • 

Thanks.