Repository navigation
MoTSR is a 2x temporal model for real-life video, released by xyether. The non-temporal MoSRv2 backbone is by umzi2; upstream notices are included.
This release includes the original EMA safetensors checkpoint, the supplied seven-input recurrent ONNX, a spatial seed ONNX, and prebuilt FP16 TensorRT 11.0.0.114 engines for NVIDIA T4. The two-cell Colab notebook downloads checksum-verified engines and uses floating-point recurrent feedback.
The supplied 52-frame preview clip was processed in the T4 runtime. H.264 2560x1440 and H.265 1920x1080 tests kept the original frame count and exact 2.12-second timing. Portrait/odd dimensions, cut resets, audio retention, PyTorch inference and ONNX export were also checked. See validation.json in the repository for numeric comparisons and environment details.
MIT applies to xyether's temporal code, notebook and weights. Upstream licenses remain included. Preview footage is separate from that license.
Results are experimental and still being developed. No universal runtime compatibility or temporal-quality benchmark is claimed.