CreativeVR: Diffusion-Prior-Guided Approach for Structure and Motion Restoration in Generative and Real Videos

Tejas Panambur, Ishan Rajendrakumar Dave, Chongjian GE, Ersin Yumer, Xue Bai

Keywords: Generative Video Refinement, Generative Video Models

Paper Track: Long Paper (archival)

Abstract: Generative video models are rapidly entering cinematic production pipelines, yet even state-of-the-art text-to-video (T2V) systems produce fine-scale structural artifacts, distorted faces and hands, warped backgrounds, and temporally inconsistent motion that prevent direct use in professional workflows. Re-generation is costly, non-deterministic, and risks deviating from creative intent, motivating a dedicated post-production refinement stage. Classical video restoration and super-resolution (VR/VSR) methods, in contrast, are tuned for synthetic degradations such as blur and downsampling and tend to sharpen these artifacts rather than repair them, while diffusion-prior restorers are usually trained on photometric noise and offer little control over the trade-off between perceptual quality and fidelity. We introduce CreativeVR, a diffusion-prior-guided video restoration framework for AI-generated (AIGC) and real videos with severe structural and temporal artifacts. Built as a lightweight deep adapter on a frozen T2V DiT backbone, CreativeVR exposes a single precision knob that lets editors and artists smoothly trade off between faithful detail preservation and aggressive structure/motion-corrective synthesis, functioning as a creative control surface in the post-production pipeline. Our key technical contribution is a temporally coherent synthetic degradation module that composes morphing, directional motion blur, and grid-based warping to simulate realistic geometric failures during training, aligning the diffusion prior toward the hard failure modes of modern generators. We evaluate on the curated AIGC54 benchmark spanning outputs from five T2V models (Veo3, Sora, Pika, Firefly, Ray3), using FIQA, perceptual, and GPT-based multi-aspect scoring. CreativeVR achieves state-of-the-art AIGC refinement quality while remaining competitive on standard video restoration benchmarks, with practical throughput (~13 FPS @ 720p on a single 80 GB A100).

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 9