Practical overview
When it fits
Use CineScale when you want to investigate or run the documented Wan2.2-based workflow for high-resolution prompt-to-video generation or prompt-guided refinement of an existing clip.
Who it is for
- Researchers studying high-resolution video generation with diffusion models.
- Developers working with Wan2.2 who want to explore prompt-based generation or video refinement at higher resolutions.
Documented uses
- Generate video from a prompt using a pretrained Wan2.2 checkpoint.
- Refine an existing video using a prompt and a higher-resolution target.
- Explore a tuning-free approach to high-resolution video diffusion.
Requirements & limits
The project is a public GitHub repository under the Apache-2.0 license. The README describes a local code workflow based on Wan2.2, using Python, model checkpoints, and distributed CUDA execution in its main example.
- The documented setup follows the original Wan2.2 repository rather than providing a standalone hosted workflow.
- The main generation example uses torchrun with five processes and CUDA devices; the README also shows a separate Wan2.2 TI2V-5B command.
- For video refinement, the README describes frame sampling, VAE encoding, latent upscaling, and high-resolution refinement.
- The project reports stronger aesthetic and imaging quality in its evaluation while subject and background consistency and motion smoothness remain comparable to the base model.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare prompts for Wan2.2 generation
Set up a prompt file for the documented CineScale text-to-video workflow.
Before you start
- CineScale and the Wan2.2 repository setup; the README says setup follows the original Wan2.2 repository.
- A prompts.json file in the Wan2.2 directory.
Steps
Define the desired video prompt in Wan2.2/prompts.json under the prompts list.
Use the documented CineScale/Wan2.2/cinescale.py workflow with the prompt-file and checkpoint arguments described in the README.
Expected outcome
A prompt file is prepared for the documented Wan2.2-based CineScale generation workflow.
Guide limitations
- The README marks the repository under construction.
- The available documentation does not establish successful execution or hardware compatibility.
Prepare a video for high-resolution refinement
Use the documented video-to-video inputs and settings to configure a refinement run.
Before you start
- CineScale with the Wan2.2 setup and a model checkpoint.
- An input video and a prompt describing the intended refinement.
Steps
Set the refinement prompt in Wan2.2/prompts.json.
Provide the source clip with the documented --input_video argument.
Set the target size with --size; the README example uses 3840*2160.
Keep --round_noise_steps below --sample_steps, as instructed in the README.
Expected outcome
A video-to-video refinement command is configured with an input clip, guidance prompt, and target resolution.
Guide limitations
- The README notes that a full restart discards the input.
- The repository is under construction; the documented example is not an execution or compatibility guarantee.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- Eyeline-Labs/CineScale · Official homepage title · retained in its source language