Real-time 30→60 fps video frame interpolation on mobile NPUs by reusing H.264 motion vectors instead of learned optical flow. 12.8 ms 1080p inference at 8-bit integer precision; 28.4 ms median end-to-end latency on Snapdragon 8 Gen 3.
- 12.8 ms
- 1080p NPU inference, 8-bit integer
- 28.4 ms
- Median end-to-end latency, Snapdragon 8 Gen 3
- 30 → 60 fps
- Real-time frame doubling on mobile
TongXing bets that the phone a blind or low-vision user already owns can be the whole assistive device, with no smart glasses or electronic cane. Worn on the chest with Bluetooth earphones, it keeps obstacle warning on the phone's own NPU so safety never waits on the network, and a local voice agent completes touch-first app tasks such as navigation and ride-hailing.
- National 2nd Prize
- 4C2026 award certificate
- 35.16 ms
- Mean obstacle-warning latency on the phone NPU
- 3 compute paths
- NPU vision, GPU LLM, and CPU offline speech on one phone
piko
630stars
An independent PikPak client I build and maintain for Android, Windows, and macOS. It is where I run extensive experiments on my own UI/UX and architectural ideas, from the clip feed to transfer and playback design. Shipping releases and working through user feedback has given me practical experience maintaining a complex multi-platform system.
open-ani/animeko
21.4kstarsAs a member of the open-ani organization, I work across Animeko and its mediamp playback runtime on cross-platform infrastructure and media pipelines: the Windows on Arm port and packaging, Linux OpenGL mpv rendering, native Windows multi-touch and pen input, ExoPlayer WSOLA audio time-stretching, and a PikPak cloud torrent engine.
- 42 merged PRs
- Across animeko, mediamp, ani-website
- +33,483 / -6,192
- Lines changed in merged PRs
- Since May 2026
- Latest merge on 2026-10-22
- feat(pikpak): run PikPak as a torrent engine alongside anitorrent+9,502 / -2,986 #3444
- feat(exoplayer): add opt-in native WSOLA audio time-stretch+2,179 / -2 #47
- feat(mpv): add Linux OpenGL playback support+1,930 / -66 #40
- feat(desktop): support multi-touch and pen input on Windows+1,268 / -356 #3461
Upstream contributions
- CMake: Enable ARM64 BCJ filters in bundled liblzma+54 / -0 !12202
- fresh: Vi-mode register behavior and Windows ARM64 releases
- pikpakcli: Download a clipped time range of a remote video+496 / -140 #64
- media-kit: Windows ARM64 libmpv for Kazumi+45 / -4 #34
- DankMaterialShell: Show WiFi connection progress+59 / -13 #2684
- llama.cpp: Diagnose CJK tool-call history pollution
Publications
- 2026 arXiv Training Code
ANVIL: Accelerator-Native Video Interpolation via Codec Motion Vector Priors
Shibo Liu
Preprint arXiv preprint
- 2026 GitHub
RAFNet: A Relation-Aware Adaptive Fusion Network for Dense Classroom Student Behavior Recognition
Yongying Xia*, Elankovan A. Sundararajan, Shibo Liu, Lei Xiao
Under review The Visual Computer