在每一次迭代加入cameratoken的必要性

[ysngfo]

您好，我看代码中似乎在每一次采样中均加入了cameratoken，但只取了最后一次的结果，为什么要这么处理而不是在模型生成结束后再加入cameratoken呢

[haoningwu3639]

Good question.
That is indeed a straightforward solution, but our motivation is to enable SceneGen’s position head to output accurate positions across arbitrary noise levels. This aims to (1) improve the robustness of position prediction and (2) encourage the generation branch to maintain geometric plausibility even in the early stages of synthesis.
Our preliminary experiments observed effectiveness with this approach, but due to limited computational resources, we did not conduct a systematic, quantitative ablation study to validate these ideas.