README.md

metadata

title: Stream-DiffVSR
emoji: 🎞️
colorFrom: red
colorTo: gray
sdk: gradio
sdk_version: 6.26.0
app_file: app.py
short_description: Low-latency streamable 4x video super-resolution
python_version: '3.12'
startup_duration_timeout: 30m

Stream-DiffVSR: Low-Latency Streamable Video Super-Resolution via Auto-Regressive Diffusion

Gradio demo for Jamichsu/Stream-DiffVSR (paper, code, project page).

Stream-DiffVSR is a causally conditioned diffusion framework for online 4× video super-resolution. It combines a four-step distilled denoiser, an Auto-regressive Temporal Guidance (ARTG) module that injects motion-aligned cues during latent denoising, and a lightweight temporal-aware decoder with a Temporal Processor Module (TPM). Every frame is restored from past frames only, so there is no look-ahead latency.

The demo mirrors the authors' reference inference.py: the official StreamDiffVSRPipeline, TemporalAutoencoderTiny and custom DDIM scheduler from the GitHub repo, RAFT-large optical flow for the auto-regressive warping, empty prompt, guidance_scale=0, 4 denoising steps.

Examples

The bundled example clips were produced by 4× bicubic downsampling of videos from linoyts/repo-to-space-example-videos.

Notes

• Input is downscaled so its longest side is at most max_input_size (default 320 px) before the 4× upscale, and the clip is truncated to num_frames (default 30) to fit in a ZeroGPU window.
• TensorRT acceleration from the original repo is not used here.

@article{shiu2025stream,
  title={Stream-DiffVSR: Low-Latency Streamable Video Super-Resolution via Auto-Regressive Diffusion},
  author={Shiu, Hau-Shiang and Lin, Chin-Yang and Wang, Zhixiang and Hsiao, Chi-Wei and Yu, Po-Fan and Chen, Yu-Chih and Liu, Yu-Lun},
  journal={arXiv preprint arXiv:2512.23709},
  year={2025}
}