FlashVSR+ ONNX
ONNX versions of a fast video upscaling model
2 min readupdated
A set of weights and graphs in ONNX format for the FlashVSR+ model, ported by the moonbox team to run on various devices via ONNX Runtime powered by CPU, CUDA, or DirectML. Based on the Wan2.1-T2V-1.3B architecture, it enables real-time streaming video upscaling with hybrid attention. The repository includes fp16 and int8 versions with ready-to-use preprocessing and inference scripts. Useful for developers and video enthusiasts looking to run local video upscaling on supported hardware.
Key features
- ONNX format support for integration into third-party apps and running via ONNX Runtime.
- Two precision options to choose from: fp16 and int8 (with 8-bit DiT weight quantization).
- Includes a ready-to-use frame processing pipeline that works similarly to the original repository.
- High frame quality preservation with minimal PSNR loss compared to the original PyTorch solution.
What you can do
- Run local video upscaling using a Python script or command line.
- Integrate ONNX graphs into your own video processing projects using CUDA or DirectML.
- Test model performance in various precision modes on your own hardware.
Example
import sys
from PIL import Image
from pathlib import Path
sys.path.insert(0, "fp16")
import pipeline
p = pipeline.FlashVSRPlus("fp16", "cuda")
frames = [Image.open(f) for f in sorted(Path("clip").iterdir())]
out = p(frames, width=1280, height=768, seed=1)
Read more
Details• fp16 precision: average PSNR 43.0 / 39.8 dB on test clips.
• int8 precision: average PSNR 42.9 / 39.6 dB on test clips.
• Hardware requirements: tested on CUDA, requires a CUDA-compatible graphics card and sufficient memory for caching.
• License: Apache-2.0.










Comments