GemmaCaption-Pipe

GemmaCaption-Pipe is a Docker-first, observation-first video captioning system designed to produce accurate, distinctive captions in multiple writing styles. Instead of sending an entire video to a model, the pipeline dynamically extracts between six and twelve uniformly distributed frames based on video duration. Every frame is labeled with its timestamp and supplied chronologically, preserving the clip’s temporal structure. Gemma analyzes these frames together and returns a compact, schema-enforced JSON evidence packet containing the setting, subjects, actions, important objects, timeline, temporal highlights, camera behavior, and high-confidence caption facts. Separating visual perception from caption writing gives every style a consistent factual foundation and makes unsupported details easier to identify. The pipeline independently generates two candidates for each requested style: formal, sarcastic, humorous technology, and humorous non-technical. Dedicated persona prompts keep the styles distinct while requiring every caption to remain recognizable and visually grounded. Prompt examples are dynamically reordered to reduce repetitive output and mechanical template copying. An independent, frame-aware judge evaluates both candidates for factual accuracy and style compliance. The weaker score determines the calibrated result, preventing strong humor from compensating for hallucinated details. Captions below the configured quality threshold are regenerated with the previous candidates and actionable judge feedback, for up to three quality rounds. GemmaCaption-Pipe supports direct OpenRouter and Fireworks access or an optional authenticated Cloudflare Worker proxy. It includes strict response schemas, deterministic policy checks, bounded concurrency, transport retries, request timeouts, safe fallbacks, Docker deployment, and secret-free configuration templates for reproducible batch processing.