Taming Camera-Controlled Video Generation with Verifiable Geometry Reward

TL;DR

This work presents CamVerse, an online reinforcement learning post-training framework that optimizes a pretrained video diffusion model for precise camera control. Central to CamVerse is a verifiable geometry reward that computes dense, segment-level alignment scores by estimating 3D camera trajectories for generated and reference videos and comparing relative poses within short segments. The method uses group-relative policy optimization (GRPO) and a scale-consistent alignment process (Umeyama) to steadily improve camera-control accuracy and geometric coherence, backed by a large-scale dataset spanning diverse camera motions. Experiments on RealEstate10K and curated data demonstrate notable gains over supervised fine-tuning baselines in camera alignment and geometry, with only modest compromises in perceptual quality. Overall, CamVerse shows that online RL with geometry-aware rewards can meaningfully advance controllable video generation beyond traditional SFT approaches.

Abstract

Recent advances in video diffusion models have remarkably improved camera-controlled video generation, but most methods rely solely on supervised fine-tuning (SFT), leaving online reinforcement learning (RL) post-training largely underexplored. In this work, we introduce an online RL post-training framework that optimizes a pretrained video generator for precise camera control. To make RL effective in this setting, we design a verifiable geometry reward that delivers dense segment-level feedback to guide model optimization. Specifically, we estimate the 3D camera trajectories for both generated and reference videos, divide each trajectory into short segments, and compute segment-wise relative poses. The reward function then compares each generated-reference segment pair and assigns an alignment score as the reward signal, which helps alleviate reward sparsity and improve optimization efficiency. Moreover, we construct a comprehensive dataset featuring diverse large-amplitude camera motions and scenes with varied subject dynamics. Extensive experiments show that our online RL post-training clearly outperforms SFT baselines across multiple aspects, including camera-control accuracy, geometric consistency, and visual quality, demonstrating its superiority in advancing camera-controlled video generation.

Figures (3)

• Figure 1: Comparison between the absolute and relative pose error. Given a reference trajectory, the model randomly generates two videos with corresponding trajectories. The absolute pose error prefers global-similar but locally incorrect results, while the relative pose error emphasizes locally consistent and smooth results. The latter is more aligned with the task characteristics.
• Figure 2: Framework of CamVerse. Given the input first frame, a text prompt, and a reference camera trajectory, the video diffusion model $\pi_\theta$ samples $G$ rollouts by a stochastic reverse process (SDE sampling). Each rollout is partitioned into non-overlapping segments. For each rollout, a large 3D model is used to estimate the camera trajectory. After aligning the generated trajectory to the reference one, the verifiable geometry reward compares each pair of generated-reference segments in a relative manner and assigns a dense alignment score. With this type of reward signal, we conduct GRPO fine-tuning, improving camera-control accuracy and geometric consistency.
• Figure 3: Illustrations of qualitative results. For each example, we condition on a reference camera trajectory (start in blue, end in red) and show uniformly sampled frames from $T=0$ to $T=5$s. The first and second rows are text-to-video (T2V) cases. The third and fourth rows are image-to-video (I2V) cases. The last row is a failure case. In the first half of the clip, the camera does not execute right-forward translation and instead moves straight forward, resulting in early trajectory misalignment.