VisionScribe: AI Video Subtitle Studio

VisionScribe is an advanced, end-to-end AI captioning and subtitle generator designed to solve a major problem for content creators, developers, and educators: transcribing software walkthroughs and demo videos. Standard audio transcription tools only capture what is spoken, missing the crucial context of what is happening on the screen, VisionScribe addresses this by analyzing the visual states of video recordings. Under the hood, VisionScribe implements a highly optimized processing pipeline: 1. FFmpeg Slicing: Slices uploaded videos into frame sequences. 2. Local Heuristic Deduplication: Compares frame file-size deltas locally. In software walkthroughs, the screen often remains static for several seconds. By detecting duplicates locally and bypassing the VLM for identical frames, the tool reduces Fireworks AI token costs by 60% to 80% and drastically speeds up execution. 3. Parallel Multimodal Analysis: Sends unique frames concurrently (5 parallel workers) to the Qwen 3.7 Plus multimodal model. 4. Scene & Script Generation: The AI groups consecutive frame descriptions into logical scenes, detects timestamps, and translates these raw screen actions into engaging, narrator-style walkthrough captions rather than dry visual summaries. 5. SRT Export: Compiles the captions into standard, time-stamped SubRip (.srt) files for immediate use. AMD Technology & Hardware Integration: Our backend routes visual and language processing to Fireworks AI's infrastructure, which is powered by high-capacity AMD Instinct™ MI300X GPUs (192GB HBM3 memory, 5.3 TB/s bandwidth), making heavy parallel VLM frame description requests finish in seconds. Locally, the video splitting leverages FFmpeg, which is compatible with AMD Advanced Media Framework (AMF). Our roadmap includes bringing the models entirely on-device using AMD ROCm™ or the AMD Ryzen™ AI NPU via ONNX Runtime for a private, offline captioning experience.