Abstract
Abstract
FFram: Command-line media tools such as FFmpeg offer unmatched capability, but they impose a steep learning curve, depend on brittle text-based telemetry, and provide no built-in hardware-failure resilience, quality awareness, or distribution. We present FFram, an open-source (MIT) Python terminal user interface that exposes a diverse, extensible catalog of operations and is built around four mechanisms: (i) direct container probing through PyAV bindings instead of ffprobe subprocesses; (ii) an auto-heuristic recovery loop that intercepts out-of-memory and GPU-encoder failures and retries with a mutated command; (iii) Auto-CRF, a binary search over the Constant Rate Factor that uses libvmaf on a 5-second sample to meet a VMAF target of 95; and (iv) a zero-dependency XML-RPC master/worker mode for chunk-parallel encoding. In our evaluation, PyAV probing is 6.51× faster than ffprobe (4.53 ms vs. 29.51 ms, n=20). AutoCRF raises VMAF from 93.16 to 96.19 on a low-motion workload (+37.4% size) and from 87.69 to 96.58 on a high-motion workload (+81.1% size), where a static CRF 23 misses the target. Simulated fault injection of two failure classes was recovered in both cases. A zero-dependency XML-RPC cluster architecture is presented, though its empirical evaluation is deferred. All experiments use synthetic workloads on one machine; we discuss the resulting limits on generalization.