Drama Series Pipeline

Video rules: no model-generated text in the image except subtitles (on-screen texts drawn in post); audio matched per shot instead of a generic track; characters stay true to their reference images; tests always use the real characters; no commercial SaaS video services; outgoing messages only as drafts; show shotlist and start frames before expensive renders.

Pipeline lessons: run only one dispatcher/render process; terminate Colab sessions after GPU work; estimate costs realistically before starting; compress videos over 200 MB for sending; the RunPod API needs a User-Agent header; use reference_image_short_edge=1024 against OOM; ElevenLabs PCM needs the Pro plan, use mp3_44100_192; long ChatGPT chats crash the browser tab, start fresh chats.

Reference run: a 14-shot, ~79 s episode cost about $7.90 ($7.04 H3 pod on RTX PRO 6000 ~170 min + $0.86 LatentSync pod), roughly $0.10 per finished second. LatentSync fails when no face is detected (keep the H3 clip); upscale/slow-mo can fall back to ffmpeg on the box CPU when no GPU is available in the gated datacenter.

Services used: RunPod (prepaid, API key as RUNPOD_API_KEY), ElevenLabs (ELEVENLABS_API_KEY), ChatGPT and optionally Suno via the box browser, optional Colab Pro as extra GPU.

Use when the owner wants a vertical (9:16) short-drama series episode (ReelShort/DramaBox style) - bible and script via ChatGPT, ElevenLabs voices, MiniMax-H3 Ref2VA shots on RunPod with the owner's likeness refs, lip-sync, upscale, grade, music at -14 LUFS, multi-language subtitles.

First conversation after importing the Drama series pipeline template - set up services, budget, series concept, cast likeness and languages, then make episode 1.