Question about Image-to-Video Comparison with CausVid and Self-Forcing (Checkpoints Request)

[jiajiaxiaoshen]

Thank you for your excellent work and for releasing the demos and code.

I have a question regarding the image-to-video (I2V) comparisons shown in the paper/demo. In some of the examples, the first frame appears to be identical (or very similar) to those produced by methods like CausVid and Self-Forcing, which makes me wonder:
1. Did you perform direct comparisons with CausVid and Self-Forcing under the same I2V setting?
2. If so, would it be possible to share the exact configs and checkpoints used for I2V when running CausVid and Self-Forcing for comparison?

Access to the specific checkpoints (or configuration details) would greatly help me.

Thank you very much for your time and clarification!

[zhuhz22]

I’m a bit confused by your question, and I hope I’m not misunderstanding. Self Forcing and CausVid do not support I2V—they only support T2V—so it’s unclear to me what comparison you are referring to.

In our paper/demos, we only implement and compare T2V. I2V is an additional feature we provide in the open-source release, and it is not part of the paper’s evaluation.

To help clarify, could you please point to the specific examples you are referring to when you say:

In some of the examples, the first frame appears to be identical (or very similar) to those produced by methods like CausVid and Self-Forcing.

Because from my understanding, we did not present any comparisons under an I2V setting.

[jiajiaxiaoshen]

Thank you very much for the clarification — that helps a lot.

I have two follow-up questions:
1. Regarding the similarity of the first frames (human):
Is this similarity controlled purely through the prompt and random seed, or are there any additional mechanisms?
2. My apologies for the confusion earlier. I asked because the Self-Forcing codebase provides an I2V interface, so I initially assumed you might have also conducted I2V experiments or comparisons.

As an additional question:
Does the ability to perform I2V generation depend on whether the model is using the chunk-wise or frame-wise variant? In other words, is I2V naturally compatible with both variants, or is it more closely related to one of them?

Thank you again for your patience and clarification.

[zhuhz22]

1. Their similarity comes from the fact that they share the same base model (Wan2.1-1.3B) and the same random seed. Since both are fine-tuned from the same underlying model, it is reasonable to expect the “overall look” of the generations to be broadly similar—but Causal Forcing still delivers a significant improvement in visual quality and motion dynamics, as you can see from the demos.

2. Although Self Forcing exposes an i2v interface in the code, they neither open-source nor claim support for I2V anywhere.

The frame-wise setting natively unifies T2V and I2V: because the generation unit is a single frame, you can simply set the first frame as the conditioning image and then autoregressively generate the subsequent frames. This is exactly the appeal of AR-diffusion. In contrast, for chunk-wise models, the generation unit spans more than one frame, so you cannot straightforwardly set the first chunk as the conditioning image; instead, I2V requires separate training.