FlashVSR Examples¶

FlashVSR performs streaming video super-resolution from a low-resolution input video and writes the restored frames to an MP4 file.

Model Source¶

Model	HuggingFace	ModelScope	Purpose
FlashVSR v1.1 BF16	lzx1413/FlashVSR-v1.1-BF16	lzx1413/FlashVSR-v1.1-BF16	Streaming video restoration checkpoint

Feature Support¶

Feature	Support	Notes
Streaming video super-resolution	Supported	Processes the input in stateful chunks
Multi-GPU inference	Supported	Ulysses sequence parallelism and FSDP through `--gpu_num`
Quantization	Unsupported	The example loads the released BF16 checkpoint
CPU offload	Unsupported	No CPU-offload option is exposed
Feature cache	N/A	Not used by the streaming restoration pipeline
Server API	Unsupported	The script is an offline CLI entry point

Requirements¶

GPU: one or more CUDA GPUs; the example has been exercised on NVIDIA RTX 5090 GPUs
Software: the standard TeleFuser installation, ffmpeg, tf-kernel, and Block-Sparse-Attention
Input assets: a readable input video

Install TeleFuser by following the development setup. Build tf-kernel with its repository Makefile and install Block-Sparse-Attention according to its upstream instructions.

Model Directory¶

/path/to/FlashVSR-v1.1-BF16/
|-- flashvsr11_dit_streaming_dmd_5dc619.safetensors
\-- TCDecoder.ckpt

The default root is ${TF_MODEL_ZOO_PATH}/FlashVSR-v1.1-BF16; pass --model_root to use another location.

Quick Start¶

python examples/flashvsr/flashvsr_stream.py \
  --input_video /path/to/input.mp4 \
  --scale 4 \
  --model_root /path/to/FlashVSR-v1.1-BF16 \
  --output work_dirs/flashvsr-restored.mp4

The command restores the input at 4x scale and writes work_dirs/flashvsr-restored.mp4.

Examples¶

Video Restoration¶

`flashvsr_stream.py`¶

Use this entry point for stateful, chunked restoration of a local video.

python examples/flashvsr/flashvsr_stream.py \
  -i /path/to/input.mp4 -s 4 --gpu_num 2 \
  --model_root /path/to/FlashVSR-v1.1-BF16 \
  -o work_dirs/flashvsr-restored-sp2.mp4

Key options:

Option	Default	Description
`-i`, `--input_video`	Required	Input low-resolution video
`-s`, `--scale`	`4`	Upscaling factor
`--height`, `--width`	Auto-detected	Optional input dimensions
`--gpu_num`	`1`	Number of sequence-parallel workers
`--model_root`	`${TF_MODEL_ZOO_PATH}/FlashVSR-v1.1-BF16`	Checkpoint directory
`-o`, `--output`	Generated name	Output MP4 path
`--seed`	`0`	Random seed

Notes¶

local_range=9 emphasizes sharp detail; local_range=11 favors temporal stability.