JoyAI Video Edit — Live (ZeroGPU)

Real-time streaming video editing from your webcam. Built on the JoyAI Video Edit DiT + streaming VAE pipeline, served on ZeroGPU.

How it works

• The heavy models (DiT 16.3B, streaming VAE, MiMo-VL text encoder) are loaded at module scope under ZeroGPU's CUDA-emulation layer.
• Each session leases a real GPU in a @spaces.GPU fork; the main process is a thin WebSocket byte-pipe (H.264 both ways, MJPEG fallback).
• The custom CUDA kernels (joyomni_ops: FP8 GEMM + fused norm/rope) ship as a prebuilt cp310 / torch-2.9.1 wheel in wheels/ — matching ZeroGPU's supported stack. Attention runs on plain cuDNN SDPA (fastest on this GPU class).

Notes

• Checkpoints are downloaded once and persisted under /data, so Space restarts skip the big downloads. The slow one-time cost is the Space startup (model load), not the first user session.
• torch.compile is inert in ZeroGPU forks; each fork instead pays a short warmup (CUDA-graph capture, KV fill), so the first 2-3 chunks are slower.
• A session runs up to ~60s of GPU time. Free ZeroGPU quota is limited — this is a small-scale demo.