.png&w=256&q=75)
🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
FrameWise — Multi-Style Video Captioning
FrameWise is a fully local video captioning agent for Track 2 of the AMD Developer Hackathon. Given a clip and a set of requested styles, it produces a faithful caption in each tone — formal, sarcastic, humorous_tech, and humorous_non_tech — without any external API. The pipeline downloads the clip, samples evenly-spaced frames, and feeds them to a video-native vision-language model running locally on GPU. Rather than captioning each style separately, it issues a single grounded call per clip that emits all requested styles as strict JSON, so every caption reflects the same observed facts — sharing the expensive vision encoding and keeping runtime well within the 10-minute budget. An optional two-pass mode separates factual grounding from styling for tougher scenes. Engineering choices target the contest's real constraints: a lean container image that stays under the size cap, weights bundled for zero network dependency at evaluation, and robust JSON parsing with per-style retries so no requested style is ever missing. A shared output-contract validator powers both an in-container self-check and a two-tier test suite — a fast, dependency-light contract test plus an opt-in end-to-end container test — ensuring the results.json the judge sees is always valid.
13 Jul 2026
