ClipForger-AMD

ClipForger-AMD is a Track 2 video captioning pipeline built for short hackathon clips of 30 to 120 seconds. The project seperates regular accessibility captions from the creative judged outputs. First the uploaded video gets processed through FFmpeg for audio extraction and keyframe analysis. The audio goes to Groq Whisper for transcription while the keyframes are analyzed by Groq vision for visual context both run in parallel to save time. The timed transcript gets exported as SRT/VTT/TXT and burned into a downloadable captioned MP4. This keeps subtitles factual and readable for viewers. After the caption track is built, ClipForger-AMD creates four different styled outputs for the same clip: formal, sarcastic, humorous-tech, and humorous-non-tech. Each style has its own dedicated LLM call with a custom system prompt and few-shot examples so there is no mixing between styles. The generation uses Gemma 4 through OpenRouter as the main model, with Groq Llama as a fallback if Gemma is not available. All four style calls run in parallel using Python threads so generating all four takes about the same time as one. Each output is grounded in the transcript and visual evidence, scored for tone and accuracy, editable in the UI, and saved into a export-ready Submission JSON. The app includes seperate Captions and Summaries tabs, export readiness checkboxes, captioned video download, provider fallback diagnostics, and Docker support for the Track 2 judged container. The pipeline also detects scene categories like sports, nature, urban, food or tech and includes audio event tags. The goal is to make video captioning both practical and entertaining captions stay accurate while summaries get personality.