WorldForge V2: Eliciting Camera-Trajectory Alignment in Video Diffusion Models

Forum topic · 小凯 · 2026-10-24

Summary

WorldForge V2 is a new framework for precise camera control in video diffusion models while preserving their pretrained generative capabilities. It introduces a training-free guidance framework that uses warped references to iteratively refine generation along a target camera path. Optical-flow analysis identifies motion-related latent channels for selective guidance, helping decouple motion control from appearance, while a dual-path strategy compares guided and unguided predictions to correct artifacts from imperfect references. To reduce refinement cost, v-cache approximates corrections using previous denoising predictions, avoiding extra model evaluations. The authors also extend motion localization from latent channels to network blocks with Trajectory LoRA, concentrating adaptation on trajectory-sensitive blocks via a two-stage training strategy. Remarkably, a single training scene suffices to generalize to unseen scenes, requiring minimal training data. Quantitative and qualitative evaluations show state-of-the-art generation quality and trajectory accuracy in camera-controlled video generation. Paper: arXiv 2610.31640.

Paper Overview

Field: Computer Vision
Authors: Chenxi Song, Yanming Yang, Tong Zhao, Ruibo Li, Joey Tianyi Zhou, Chi Zhang
Published: 2026-10-22
arXiv: 2610.31640

Summary

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 like task-specific fine-tuning and post-hoc warping often introduce visual artifacts, generalize poorly, or incur high computational costs.

WorldForge V2 achieves accurate camera control while preserving the generative capabilities of pretrained video models:

Training-free guidance framework: Uses warped references to iteratively refine generation along a target camera path.
Selective guidance via optical flow: Optical-flow analysis identifies motion-related latent channels, helping decouple motion control from appearance.
Dual-path strategy: Compares guided and unguided predictions to correct artifacts caused by imperfect references.
v-cache: Approximates corrections using previous denoising predictions, avoiding additional model evaluations and reducing refinement cost.
Trajectory LoRA: Extends motion localization from latent channels to network blocks, concentrating adaptation on the blocks most sensitive to trajectory guidance via a two-stage training strategy. Reliable camera control is achieved with very little training data — a single training scene is enough to generalize to unseen scenes.

Quantitative and qualitative evaluations show state-of-the-art generation quality and trajectory accuracy for camera-controlled video generation.

Original Abstract (excerpt)

> 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. We propose WorldForge V2 to achieve accurate camera control while preserving the generative capabilities of pretrained video models. We develop a training-free guidance framework using warped references to iteratively refine generation along a target camera path. Optical-flow analysis identifies motion-related latent channels for selective guidance. A dual-path strategy compares guided and unguided predictions to correct artifacts...

--- *Auto-collected on 2026-10-24*

WorldForge V2: Eliciting Camera-Trajectory Alignment in Video Diffusion Models

Summary

Paper Overview

Summary

Original Abstract (excerpt)

Tags