PyTorch Consolidates Media Stack: TorchCodec Becomes Single Home for Decode/Encode

PyTorch · x · 2026-10-06

Meta published a technical debrief on how PyTorch's media stack has been reorganized over the past two years:

• TorchCodec is now the single home for decoding/encoding images, video, and audio on both CPU and CUDA. Previously, capabilities were scattered and duplicated: TorchVision had multiple entry points across three backends (only PyAV worked out of the box), while TorchAudio had StreamReader/StreamWriter and multiple audio backends.
• TorchVision and TorchAudio now focus strictly on high-performance media transforms for images/video and audio, respectively.
• Models and datasets no longer live in these libraries — the team points users to the wider ecosystem like HuggingFace.
• The motivation: media is now central to model development, from VLM training to generative models that must decode inputs and encode outputs.

Original post →

More from Infra

• llama.cpp beats Ollama by ~30% prompt eval speed on RTX 5060 Ti with Clef Flash 9B — ngxson · 2026-10-06
• Clef Flash 9B Q8_0 on RTX 5060 Ti: llama.cpp ~30% faster prompt eval than Ollama (3,000 vs 2,100 t/s) — ngxson · 2026-10-06
• OpenRouter hits 2T tokens a day as DeepSeek v4.1 traffic surges 10x overnight — stuffyokodraws · 2026-10-06
• SpaceX files for 32.4-mile Florida gas pipeline to make Starship methane on-site — DimaZeniuk · 2026-10-06
• Inference Companies Ship Gateways, Gateway Companies Ship Inference — the Lines Are Blurring — michellechen · 2026-10-06
• Can an 8GB Radeon 5700 Run Image-to-Video? ComfyUI Low-VRAM Question — Fluffy-Composer9675 · 2026-10-06

Infra channel →