About the pixel l2 loss of stage 3

Thank you very much for your impressive work. I have been carefully reading the paper and found the design of Stage 3 particularly interesting.

My current understanding is as follows: suppose the two selected latents correspond to temporal indices
[𝑥,𝑥+1].
During decoding, the VAE actually decodes the latent sequence [0,...,x+1 ]to preserve temporal causality, while the latents [0,x-1] are detached from gradients. Then, pixel-domain supervision is applied only to the decoded frames corresponding to the last few frames (e.g., the last 8 frames).

I would like to confirm whether this interpretation is correct, or if instead the decoding is performed only on the selected latents [𝑥,𝑥+1] without explicitly decoding earlier frames.

作者您好，
感谢您的工作为社区带来的贡献，我在尝试应用训练方法到自有数据时，对于阶段3的实现有一点疑问。
在阶段3， 论文中提到具体在pixel domain参与监督的只有随机选择的两个latents，我想知道具体是怎么选择的，
比如选择了时间维度上[x,x+1]的latents的话，做法是否是用VAE decode [0,x+1]的latents，其中[0,x-1]需要detach处理，然后在最后8帧上进行像素域的监督？

No file, test, or entry point is named in the issue. Read the Stage 3 training implementation and its VAE decoding path first, then document whether earlier latents are decoded and detached and which selected frames receive pixel-domain supervision.