Causal Forcing FAQ
输入“/”快速插入内容
Causal Forcing FAQ
Causal Forcing FAQ
用户3747用户37475月12日修改
中文版在下方
1.
What is the relationship between Causal Forcing and Self Forcing?
•
Causal Forcing modifies the ODE initialization stage of Self Forcing, addressing the architectural gap from a bidirectional model to an AR model.
2.
Does Causal Forcing still need DMD?
•
Yes. The DMD stage in Causal Forcing is exactly the same as in Self Forcing: the student is conditioned on its self-generated history.
3.
How does Causal Forcing work? What are the stages?
•
Stage 1: Train a multi-step autoregressive diffusion model with teacher forcing.
•
Stage 2: Use the AR diffusion model from Stage 1 to construct causal ODE data, then train a few-step AR model.
•
Stage 3: Initialize from the Stage 2 model and train Self Forcing DMD.
4.
Is Causal Forcing much more complicated and costly than Self Forcing?
•
No. The Self Forcing pipeline is:
◦
Stage 1: Construct ODE data using a bidirectional teacher.
◦
Stage 2: DMD.
In other words, Self Forcing also requires constructing ODE data. Causal Forcing only adds one extra stage: "Stage 1: train a multi-step autoregressive diffusion model with teacher forcing."
See the following figure for a comparison of the pipelines:
5.
Does Causal Forcing bring a large improvement?
•
Yes. For example, one user reported that Self Forcing caused abrupt changes between video chunks in their task, while replacing the initialization with Causal Forcing fixed the issue. See the last few comments in https://github.com/guandeh17/Self-Forcing/issues/64, where a user provides comparison demos on Wan 2.2 5B. The improvement is even larger in the frame-wise setting.
6.
Constructing causal ODE data is time-consuming and storage-intensive. Can we avoid constructing ODE data?
•
Yes. Causal consistency distillation can be used as a replacement, and the repository already provides complete code. That's what we call Causal Forcing++.
7.
Is Stage 2 necessary? Can we directly use multi-step AR diffusion as the DMD initialization?
•
It is better than the incorrect ODE initialization used in Self Forcing, but still worse than Causal Forcing. If you want a cheaper setup, even training with a few hundred ODE samples for about 1k steps can already bring improvements.
8.
Since Causal Forcing and Self Forcing use autoregressive architectures, do they natively support long-video generation?
•
No. Autoregression does not automatically imply native long-video support. Just as autoregressive LLMs still need methods such as StreamingLLM for length extrapolation, Causal Forcing and Self Forcing also require additional modifications to generate videos beyond the training length, such as LongLive or Infinity Forcing.
More discussion:
Why use bidirectional teacher in the DMD stage ?
•
Q: In the DMD stage, do you still use a bidirectional teacher? Why not an AR teacher?
•
A: Yes. DMD only requires the student to match the teacher’s final distribution, not the generation trajectory, so a bidirectional teacher is fine. Also, bidirectional diffusion models are typically stronger than AR diffusion, so they make a better teacher.
•
Q: Then why must the ODE (or Consistency Distillation) stage use an AR teacher?
•
A: Because ODE/CD requires the student and teacher to follow the same trajectory, so their structures must be matched; an AR student cannot be trajectory-aligned with a bidirectional teacher.
ODE initialization or multi-step AR diffusion initialization ?
•
Q: Which is better as initialization: a “proper” ODE initialization or directly using multi-step AR diffusion?