VBench evaluation

[smilebetterworld]

Hi， training dataset prompt is like "Camera move right, pull out, then pan right. A whale breaching the ocean surface." which includes camera movement, but when performing VBench evaluation, the prompt would not include camera movement, just like "a drone flying over a snowy forest". So I do not understand why the model perform well in Vbench, because the model lose camera movement info and noise wrapper should not work

[lhmd]

Thanks for the question. Your understanding is correct that, for a plain VBench prompt such as “a drone flying over a snowy forest”, there is usually no explicit camera-motion phrase to parse, so the camera-trajectory based noise wrapping is not activated at inference/evaluation time.

In World-R1, noise wrapping is mainly used during RL rollouts when the training prompts contain camera instructions. It provides a trajectory prior so that the generated rollouts can be evaluated by the 3D-aware rewards, especially the trajectory-alignment term. The learned LoRA/model parameters are then updated by these rewards. Therefore, the improvement on VBench does not mean that VBench prompts secretly contain camera movement, or that noise wrapping is always applied during VBench evaluation. Rather, it reflects that the post-trained model has internalized better geometric consistency and visual regularity from the RL training process.

There are two additional points:

1. The 3D-aware reward is not only the trajectory score. It also includes meta-view and reconstruction consistency terms, while the general generation reward preserves visual/aesthetic quality. These learned effects can transfer to camera-free prompts.
2. The periodic dynamic-only training stage uses dynamic prompts and disables the 3D reward, optimizing only the general reward. This is intended to prevent the model from becoming overly static/rigid and helps maintain motion diversity, which is relevant to VBench metrics.

So for VBench-style prompts, it is fair to say the inference-time noise wrapper may not work or may not be used. The VBench gain should be interpreted as a result of RL fine-tuning with camera-aware rollouts and 3D/general rewards, not as direct camera conditioning on the VBench prompts.