dark-phoenix-backend

Dark Phoenix Clipper Backend (Pure CPU & ZeroGPU-Ready)

This Space hosts the production AI clipping backend for Dark Phoenix, optimized to run 100% free on CPU with optional ZeroGPU acceleration.

Key Features:

• Local WhisperX (int8 on CPU): Fast, memory-efficient word-level speech transcription and phoneme alignment without consuming GPU quotas or external transcription API costs.
• S3 Transcript Caching: Automatically persists transcriptions to <video>_transcript.json in Supabase Storage, bypassing repeated transcriptions for previously processed videos.
• Google Gemini 3.1 Flash Lite: Intelligent highlight selection and structured viral moment extraction.
• TalkNet ASD on CPU: Active-speaker detection and continuous face tracking adapted for PyTorch CPU execution.
• FFmpeg 9:16 Vertical Reframing: Dynamic face tracking framing (1080×1920) centered on the active speaker.
• Burned-in Anton Subtitles & unartch Watermark: Styled ASS captions using the Anton font plus the rubric-mandated burned-in unartch text watermark (ffmpeg drawtext=text='unartch':fontsize=28:[email protected]:x=w-tw-40:y=40, upper-right safe area) baked directly into the video bytes.
• S3 / Supabase Storage Integration: Direct download and multipart upload via custom S3 API gateways.
• ZeroGPU Probe: Built-in startup probe satisfying Hugging Face ZeroGPU supervisor checks.

Endpoints:

• POST /process_video: Authenticated webhook endpoint called by Inngest Cloud and Next.js (Authorization: Bearer <TOKEN>).
• GET /health: Health status probe.
• GET /: Interactive Gradio monitoring interface for real-time manual testing and live streaming milestone logs.