Note: This repository is under construction.

Gordon Chen^†, Haonan Qiu^†, Ning Yu^*, Ziqi Huang, Paul Debevec, Ziwei Liu^*

^† Equal contribution    ^* Corresponding authors

From Nanyang Technological University and Netflix Eyeline Studios.

Most video generators are trained at limited spatial resolutions due to the scarcity of high-resolution 4K video data and the prohibitive computational cost of large-scale training on such data. Most video diffusion models are trained on 720p videos and are therefore effectively limited to generating videos at similar resolutions during inference. To address this gap, we propose CineScale. CineScale, to the best of our knowledge, is the first tuning-free inference framework enabling pretrained video diffusion models to generate high-fidelity videos at resolutions far beyond those encountered during training, without any fine-tuning.

CineScale unlocks tuning-free 4K video generation (Watch our Video Demo here)!

Qualitative 4K video generation results produced by CineScale.

Qualitative comparison with existing video generation models.

CineScale substantially improves perceptual quality at high resolution, with the strongest gains in aesthetic and imaging quality. Subject consistency, background consistency, and motion smoothness remain comparable to the base model because CineScale focuses on recovering fine-grained spatial details while preserving the temporal behavior and semantic structure established by the low-resolution generation.

VBench comparison across different target resolutions. SC: Subject Consistency; BC: Background Consistency; TF: Temporal Flickering; AQ: Aesthetic Quality; IQ: Imaging Quality. Higher is better for every metric. The best, second-best, and third-best results are marked in bold, underline, and italics, respectively.

The setup follows exactly from the original Wan2.2 Repository. (https://github.com/Wan-Video/Wan2.2)

Deffine your prompts in:

Wan2.2/prompts.json

For instance:

{
    "prompts": [
        "A little girl, lost in the city and separated from her parents in New York's Times Square, looks up. **The camera tilts up**, following her gaze. Starting from the ground, it slowly reveals the massive, glittering, and dizzying skyscrapers and billboards, powerfully emphasizing her smallness and helplessness in a vast world."
    ]   
}

and then run:

CUDA_VISIBLE_DEVICES=0,1,2,3,4 \
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
torchrun --standalone --nproc_per_node=5 CineScale/Wan2.2/cinescale.py \
  --prompts_json CineScale/Wan2.2/prompts_2.json \
  --output_dir CineScale/example_videos \
  --ckpt_dir Wan2.2-T2V-A14B \
  --frame_num 41 \
  --round_noise_steps 25 \
  --ulysses_size 5 \
  --dit_fsdp \
  --t5_cpu \
  --offload_model true 

Add --input_video to refine an existing video. The prompt in prompts.json guides the refinement.

CUDA_VISIBLE_DEVICES=0,1,2,3,4 \
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
torchrun --standalone --nproc_per_node=5 CineScale/Wan2.2/cinescale.py \
  --input_video path/to/input.mp4 \
  --prompts_json CineScale/Wan2.2/prompts.json \
  --output_dir CineScale/v2v_latents \
  --ckpt_dir Wan2.2-T2V-A14B \
  --size '3840*2160' \
  --frame_num 41 \
  --round_noise_steps 25 \
  --ulysses_size 5 \
  --dit_fsdp \
  --t5_cpu \
  --offload_model true

CineScale samples up to --frame_num frames, VAE-encodes the video, bilinearly upsamples its latent to --size, and runs high-resolution refinement. Keep --round_noise_steps below --sample_steps; a full restart discards the input.

Wan2.2 TI2V-5B checkpoints are also supported. The model variant is detected from the standard checkpoint layout, or it can be selected explicitly:

PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
python CineScale/Wan2.2/cinescale.py \
  --prompts_json CineScale/Wan2.2/prompts.json \
  --output_dir CineScale/example_videos_5b \
  --ckpt_dir Wan2.2-TI2V-5B \
  --model_variant ti2v-5B \
  --offload_model true \
  --convert_model_dtype \
  --t5_cpu

The 5B path uses its native 1280*704 prompt base, Wan2.2 VAE, 24 FPS, 50 sampling steps, and shift 5.0 unless those settings are overridden.

Wan2.1 T2V-1.3B checkpoints are supported through the single-DiT path. Its standard checkpoint layout can be detected automatically, or selected explicitly. It generates its prompt base at 832*480 while using the same 80-by-45 maximum-reference grid as Wan2.2 A14B, with 21-by-12 query tiles, horizontal/vertical thresholds of 16/9, and maximum zero-based relative offsets x=79 and y=44:

CUDA_VISIBLE_DEVICES=0,1,2,3 \
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
torchrun --standalone --nproc_per_node=4 CineScale/Wan2.2/cinescale.py \
  --prompts_json CineScale/Wan2.2/prompts.json \
  --output_dir CineScale/example_videos_1_3b \
  --ckpt_dir Wan2.1-T2V-1.3B \
  --model_variant wan2.1-t2v-1.3B \
  --block_tiled_self_attn true \
  --frame_num 41 \
  --round_noise_steps 40 \
  --ulysses_size 4 \
  --dit_fsdp \
  --t5_cpu \
  --offload_model true

This path uses an 832*480 prompt base, the Wan2.1 VAE, a single DiT, 16 FPS, 50 sampling steps, and shift 8.0 unless overridden. When decoding its high-resolution latent payloads, CineScale automatically uses 64*64 VAE latent tiles in both spatial dimensions. The tile size can be overridden with --vae_decode_tile_height and --vae_decode_tile_width.

To decode the video, run:

PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
python CineScale/Wan2.2/cinescale.py \
  --decode_latent path/to/video_latent.pt \
  --ckpt_dir checkpoint/to/model_weights/Wan2.2-T2V-A14B

To inspect, batch-decode, and optionally score every completed latent payload, use vbench_batch.py. The VAE is loaded once and the decoded MP4 files are written to a separate directory:

CUDA_VISIBLE_DEVICES=0 \
python CineScale/Wan2.2/vbench_batch.py \
  --latents_dir CineScale/example_videos_1_3b \
  --videos_dir CineScale/vbench_videos_1_3b \
  --prompts_json CineScale/Wan2.2/prompts.json \
  --ckpt_dir Wan2.1-T2V-1.3B \
  --model_variant wan2.1-t2v-1.3B \
  --latent_key final_latent \
  --run_vbench

The script prints the stored prompt for every .pt, skips existing MP4 files unless --overwrite is supplied, writes a JSON manifest beside the video directory, and runs the six dimensions supported by VBench custom-input mode. Use --list_only to inspect payloads without decoding them.

This codebase is built on top of the open-source implementation of Wan2.2 repository.

If you find CineScale useful in your research or projects, consider citing our paper:

@article{chen2026cinescale,
  title={CineScale: Tuning-Free High-Resolution Video Generation},
  author={Chen, Gordon and Qiu, Haonan and Yu, Ning and Huang, Ziqi and Debevec, Paul and Liu, Ziwei},
  journal={arXiv preprint arXiv:2508.15774},
  year={2026}
}