Video diffusion models have learned rich 3D/4D priors, making them promising for scene generation and video re-rendering. However, their use in spatial tasks is limited by imprecise camera control, spatial and temporal inconsistencies, and entangled camera motion with scene dynamics. Existing approaches, such as task-specific fine-tuning and post-process warping, often introduce visual artifacts, generalize poorly, or incur high computational costs.
Training-free guidance with warped references for iterative refinement
WorldForge V2 achieves accurate camera control while preserving the generative capabilities of pretrained video models. The method first develops a training-free guidance framework using warped references to iteratively refine generation along a target camera path. At each refinement step, the model compares current predictions against reference frames warped to approximate the desired viewpoint, allowing gradual correction of camera position and orientation without modifying the underlying diffusion model.
Optical-flow analysis for motion-channel selection
Optical-flow analysis identifies motion-related latent channels for selective guidance, helping separate motion control from appearance. By analyzing flow patterns across video frames, the method determines which latent dimensions primarily encode camera movement versus scene content. This separation allows the guidance mechanism to act on motion-specific channels while leaving appearance-related channels undisturbed, reducing the risk of distorting object identities or scene details during camera movement.
Dual-path artifact correction strategy
A dual-path strategy further compares guided and unguided predictions to correct artifacts from imperfect references. One path proceeds with camera-guided refinement using warped references, while the other maintains the original unguided generation. The comparison between these two paths reveals discrepancies introduced by less-than-perfect warp estimates, and the method adjusts the final output to retain the strengths of both approaches. This complementary handling of guided and unguided streams preserves scene appearance and object motion even when reference warps are not perfectly accurate.
v-cache for cost-effective refinement
To reduce refinement costs, v-cache approximates the correction using the preceding denoising prediction, avoiding an additional model evaluation. When refining along a camera path, each new step can reuse the latent state from the previous timestep, modified by the estimated warp. This caching technique means that camera-controlled generation does not require a full forward pass at every step, making the method significantly more efficient for long trajectories or real-time applications.
Motion localization and Trajectory LoRA
The authors further extend motion localization from latent channels to network blocks and introduce Trajectory LoRA, which concentrates adaptation on blocks most responsive to trajectory guidance. Rather than applying full fine-tuning to the entire model, Trajectory LoRA inserts low-rank adaptation modules targeted at specific transformer blocks that show the strongest response to camera trajectory inputs. A two-stage training strategy teaches these adapters to denoise the states encountered during guided generation, using limited supervision effectively. This enables reliable camera control with very little training data, with a single training scene already supporting generalization to unseen scenes.
Quantitative and qualitative evaluation results
Quantitative and qualitative evaluations across multiple datasets show that the method achieves state-of-the-art generation quality and trajectory accuracy in camera-controlled video generation. The approach outperforms existing methods on metrics measuring both visual fidelity of generated content and precision of camera path following. Ablation studies confirm that each component—warped-reference guidance, optical-flow channel selection, dual-path correction, v-cache, and Trajectory LoRA—contributes meaningfully to the overall improvement. The results demonstrate that accurate camera control in video diffusion models is achievable without sacrificing the generative quality that makes these models attractive for creative applications.