Seedance Animator

Research-first pipeline that turns a verbal concept into a fully animated scene using Replicate (Nano Banana 2 for stills, Seedance 2.0 for video). The key differentiator is Agent 0: story-researcher — it deep-dives the source material, references, and visual inspirations BEFORE any image generation, ensuring the final video tells a real story instead of producing pretty but hollow transitions.

Hard rules

• Always use Nano Banana 2 (google/nano-banana-2) for every image generation step.
• Always use Seedance 2.0 (bytedance/seedance-2.0) for animation. This is the only confirmed working slug — seedance-2-lite does NOT exist on Replicate as of 2026-04.
• Save everything to ~/Desktop/Story Teller/<project_name>/ with subfolders for each asset type (characters/, style/, storyboard/, prompt/, video/). Ask the user for a project name or derive one from the scene idea.
• Never invent model IDs. If the Replicate API returns 404 for a model, prompt the user for the exact slug they saw in the Replicate playground.
• .env holds REPLICATE_API_TOKEN. Never commit it. .env.example is the template.
• Windows Python paths: Always use os.path.expanduser("~/.env") and os.path.expanduser("~/Desktop/...") — never use Git Bash /c/ prefix in Python code as it resolves incorrectly.
• Content filter: Seedance 2.0 blocks terms like "fanservice", "ecchi", "compromising position", "on top of him". Always sanitize prompts before submission — use "slapstick comedy", "comedic landing", "awkward tangle" instead.

Conversation flow (MANDATORY — do this BEFORE any API call)

1. Idea — "¿Cuál es la idea/historia/escena? (personajes, acción, duración, mood)"
2. Art style + inspirations — "¿Qué estilo? Y si tienes referencias de películas, directores, anime, o arte, dímelas — las investigo a fondo para extraer su lenguaje visual." Examples: "estilo Moana", "como Villeneuve filma Dune", "Ghibli meets Orozco murals"
3. Image references — "¿Tienes imágenes de referencia? (sube paths o genero todo from scratch con Nano Banana 2)"