Description
What I'm building An AI tool that takes a narration audio file + a folder of images and automatically outputs a video where the right image is shown for what's being said at each moment — with smooth zoom in/out and slide in/out motion (Ken Burns style), synced to the audio. Fully automatic, no manual scripting. A working prototype already exists and handles: audio-to-cut timing (cuts land in speech pauses, never mid-word), pacing (each image held for the full point), the zoom/slide animation, and rendering. So you're building on a foundation, not from scratch. The focus of this job — the core feature: Accurate, fully-automatic matching of images to narration — "the right picture for the right words," no random images. This means: • transcribe + understand the story from the audio (Whisper) • read each image with a vision-language model (Qwen2-VL / LLaVA / GPT-4o) so matches are by meaning, not shape/color similarity (CLIP alone is not enough) • match each narration moment to the correct image How we'll work: Start with a small paid trial — get accurate matching working on one 2–3 minute test scene first. If it works, we continue with the full feature. (Performance for long videos and server deployment can be follow-up phases.) Deliverables: working matching module integrated with the existing prototype + source code + a short setup note. To apply, please answer: Have you built with FFmpeg and vision-language models? Share examples/links. In 2–3 sentences: how would you appro