The full pipeline for producing a wide-cam AI podcast video — two AI hosts in a locked-off lounge 2-shot, ~60 seconds, 16:9, dialogue-only, native audio.
When to invoke
The user wants any of:
- An AI-hosted short-form ad or VSL with two characters in a podcast-style conversation
- A promotional clip for a brand, course, community, or product
- A UGC explainer with two talking heads
- A talking-head podcast clip in 16:9 for YouTube / X / Reels (16:9 master, derive verticals after)
Required environment variables
Set these in the user's shell or .env before any pipeline step runs.
WAVESPEED_API_KEY— required. Used for both image (gpt-image-2) and video (bytedance/seedance-2.0) generation, plus the CDN upload for reference images.GEMINI_API_KEY— optional. Only needed if you want 4K starting frames via Nano Banana 2 instead of gpt-image-2's 2K.VIDJUTSU_API_KEY(orvidjutsu authtoken already saved at~/.vidjutsu/config.json) — optional. Only needed for the compliance scan and disclaimer burn.
The pipeline (7 steps)
Read each recipe page in order. Every recipe is a self-contained markdown file under recipes/.
- Define characters →
recipes/01-define-characters.md— one JSON per host - Generate starting frames (per character) →
recipes/02-generate-starting-frames.md— Wavespeed gpt-image-2 - Define the wide scene →
recipes/03-define-scene-wide.md— one JSON describing the 2-shot - Define clips (one per dialogue block) →
recipes/04-define-clips.md— one JSON per ~15s block - Generate clips →
recipes/05-generate-clips.md— Wavespeed Seedance 2.0 text-to-video, wide as reference image - Concat →
recipes/06-concat.md— ffmpeg stream-copy concatenation - Compliance scan (optional) →
recipes/07-compliance.md— VidJutsu/v1/compliance/video, plus disclaimer burn if needed
Working directory layout
Set up the project's working directory exactly as described in references/folder-structure.md before running any step. Every recipe assumes that layout.
Schema and endpoint references
When you need the exact JSON shape or API contract:
references/folder-structure.md— canonical project layoutreferences/character-spec.md— character JSON schemareferences/scene-spec.md— scene JSON schema (multi-person)references/clip-spec.md— Seedance clip JSON schemareferences/wavespeed-gpt-image-2.md— image generation endpoint, params, gotchasreferences/wavespeed-seedance-2.md— video generation endpoint, params, render timesreferences/gemini-nano-banana-2.md— alt 4K image generation endpointreferences/vidjutsu-compliance.md—/v1/compliance/videoendpoint and response shapereferences/vidjutsu-disclaimer.md—/v1/disclaimerendpoint for fine-print burn-inreferences/ffmpeg-concat.md— stream-copy concat recipe
Cost and time expectations
- One starting frame: ~30 seconds to ~2 minutes per render at 2K
- One ~15s Seedance clip: ~5–8 minutes per render at 720p; runs in parallel
- Full 4-block ~60s ad, end to end, set up: under 2 hours
- Total stack cost per finished ad: single-digit dollars
Style guardrails baked into every recipe
- 16:9 throughout (vertical or square crops are derived after, not generated separately)
- Native audio in the clip model — do not stitch voiceover on top
- No captions, overlays, music, or third-party logos in the source generation. Burn captions at the very end if needed
- Cast the two hosts with visibly different silhouettes (hair, wardrobe color, pose) so cuts read instantly
- One simple question per dialogue block, asked by Host A; Host B teaches/answers
- Casual, non-technical tone — assume the viewer is a non-builder