Research map / Video / Video generation
DynamicScaler: Seamless and Scalable Video Generation for Panoramic Scenes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DynamicScaler, a tuning-free denoising framework that uses shifting windows, global motion guidance, and panoramic projection to generate high-resolution, arbitrary-aspect-ratio, 360° and long/loopable dynamic panoramic videos from a fixed-resolution video diffusion model.
Problem
Video diffusion models are limited to fixed resolution and aspect ratio, making scene-level dynamic panoramic (including 360° equirectangular) video generation hard while keeping motion coherent and memory bounded.
Approach
Offset Shifting Denoiser shifts denoising windows vertically and horizontally (with wrap-around horizontally) at every step so that overlap regions synchronize content across windows; Global Motion Guidance generates a low-resolution video first and re-noises its upscaled version to initialize high-resolution denoising; panoramic projecting denoise applies the same shifting to perspective viewport
Key results
- Table 1: DynamicScaler reaches dynamic degree 0.783 and Q-Align(V) 0.613 versus 0.412 and 0.532 for 360DVD.
- Table 3: the full method attains image quality 0.587 and temporal flickering 0.985, while direct inference at target resolution runs out of memory.
- User study with 20 participants (Table 2) rates DynamicScaler highest in most criteria, e.g. 4.6 vs 3.3 graphics quality against 360DVD in same-case comparison.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Offset Shifting Denoising (ablation) | DynamicScaler · Ablation study | Image Quality: 0.564 -> 0.587 | Helped |
| Add Offset Shifting Denoising (ablation) | DynamicScaler · Ablation study | Dynamic Degree: 0.749 -> 0.778 | Helped |
| Add Offset Shifting Denoising (ablation) | DynamicScaler · Ablation study | Motion Smoothness: 0.948 -> 0.967 | Helped |
| Add Offset Shifting Denoising (ablation) | DynamicScaler · Ablation study | Temporal Flickering: 0.905 -> 0.985 | Helped |
| Add Offset Shifting Denoising (ablation) | DynamicScaler · Ablation study | Q-Align(V): 0.595 -> 0.616 | Helped |
| Add Global Motion Guidance (ablation) | DynamicScaler · Ablation study | Image Quality: 0.571 -> 0.587 | Helped |
| Add Global Motion Guidance (ablation) | DynamicScaler · Ablation study | Dynamic Degree: 0.765 -> 0.778 | Helped |
| Add Global Motion Guidance (ablation) | DynamicScaler · Ablation study | Motion Smoothness: 0.961 -> 0.967 | Helped |
| Add Global Motion Guidance (ablation) | DynamicScaler · Ablation study | Temporal Flickering: 0.946 -> 0.985 | Helped |
| Add Global Motion Guidance (ablation) | DynamicScaler · Ablation study | Q-Align(V): 0.598 -> 0.616 | Helped |
Limitations
- Relies on comparison with only one publicly available dynamic panorama baseline (360DVD) for quantitative results.
- User study uses 20 participants and random-case comparisons for several methods due to limited availability of baselines.
Experimental setup
Benchmarks: VBench metrics CogVideoX-style metrics
Metrics: CLIP-Score image quality dynamic degree motion smoothness temporal flickering Q-Align(I) Q-Align(V) user preference scores
Baselines: 360DVD 4K4DGen ScaleCrafter VividDream
Builds on: MultiDiffusion SyncDiffusion VideoCrafter2 VideoCrafter1 360DVD
Closest papers by meaning
- 4K4DGen: Panoramic 4D Generation at 4K Resolution (ICLR 2025 · similarity 0.91)
- PanoDiT: Panoramic Videos Generation with Diffusion Transformer (AAAI 2025 · similarity 0.90)
- ViewPoint: Panoramic Video Generation with Pretrained Diffusion Models (NeurIPS 2025 · similarity 0.89)
- PanoWan: Lifting Diffusion Video Generation Models to 360$^\circ$ with Latitude/Longitude-aware Mechanisms (NeurIPS 2025 · similarity 0.88)
- PanoWorld: Geometry-Consistent Panoramic Video World Modeling (NeurIPS 2026 · similarity 0.88)
- DiffPano: Scalable and Consistent Text to Panorama Generation with Spherical Epipolar-Aware Diffusion (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.