CineScale: Tuning-Free Framework Lets Open-Source Video Models Generate 4K at Inference
liuziwei7 · x · 2026-09-14
Researchers from NTU and Netflix Eyeline Studios released CineScale, claimed to be the first tuning-free inference framework enabling pretrained video diffusion models to generate high-fidelity video well beyond their training resolution.
- Most video models are trained at 720p and can't exceed that at inference due to scarce 4K data and training cost.
- CineScale partitions query tokens into spatial tiles while keeping access to all global keys/values, paired with Adaptively Rectified RoPE (AR-RoPE) to preserve local geometry.
- Open-source models like Wan can now output 4K videos without any fine-tuning; the method is integrated into Wan.
- Paper includes baselines against RoPE/NTK-RoPE and cascading-step ablations.
More from Multimodal
- HyperFrames + HeyGen: decent video from a basic prompt in ~10 minutes — toolstelegraph · 2026-10-30
- Hailuo MCP + Claude: chaining motion graphics generation through an agent — aziz4ai · 2026-10-30
- AI short film Nexus ep 7 cost $28,775 to make, creator shares full breakdown — NightsRadiant · 2026-10-30
- LayerShop: open-source Photoshop-style editor with ComfyUI local AI on layers — HonestTicket6013 · 2026-10-30
- AI video prompt goes viral: rally car racing a collapsing volcanic mountainside — umesh_ai · 2026-10-30
- Redditor finishes a short dinosaur movie made with Veo 3 and Seedance — HeadOpen4823 · 2026-10-30