Our take
- Activity
- 35
- Community
- 35
- Issues
- 36
- Pull requests
- 20
+2 stars in the last 30 days
- Slower to close issues than 93% of the projects we track
Project history
Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.
Stars
220Forks
6Change since Nov 14, 2026: 0 stars·0 forks
Overview
CineScale applies a tuning-free inference approach to pretrained video diffusion models, aiming to generate detailed video at resolutions beyond those used in training. It addresses the cost and data scarcity that make training video models directly at 4K difficult.
The project builds on Wan model implementations and focuses on adding spatial detail while retaining the base model’s temporal behavior and semantics. It may suit teams exploring high-resolution generation or refinement, provided they have the necessary model checkpoints and compute.
Key Features
- Extends pretrained video diffusion inference to higher target resolutions without fine-tuning.
- Supports text-to-video generation and video-to-video refinement.
- Documents Wan2.2 A14B and TI2V-5B checkpoint paths, as well as Wan2.1 T2V-1.3B.
- Provides latent decoding and batch video decoding utilities.
- Includes optional VBench scoring in the batch workflow.
- Exposes settings for frame count, resolution, sampling, distributed execution, and model offloading.
Use Cases
- Video-generation researchers testing high-resolution inference without retraining a base model.
- Studios and creators refining existing clips toward larger output sizes, where supported checkpoints and GPU resources are available.
- Practitioners comparing high-resolution output quality using the project’s reported VBench dimensions.
- Developers evaluating Wan-based generation workflows or decoding saved latent outputs.
Getting Started
Follow the setup for the Wan2.2 repository, then use the CineScale README for checkpoint setup and complete commands. The documented generation workflow uses torchrun with CineScale/Wan2.2/cinescale.py and a Wan checkpoint directory. See the project page for demonstrations.
Considerations
- The repository README says the project is under construction, so expect implementation and documentation to evolve.
- Generation requires compatible Wan checkpoints and substantial GPU resources; examples use multi-GPU execution and include memory-management options.
- Output resolution and supported model variants depend on the selected checkpoint and configuration. Consult the README before planning a run.
- CineScale focuses on spatial detail. The README reports that temporal and consistency metrics remain comparable to the base model, rather than claiming universal improvements across all dimensions.
Found this useful?
Share it with someone who would like CineScale.