storycast-seed-agents

Submission Package

Updated: April 18, 2026

Submission Links

• Landing page: https://rohan5commit.github.io/storycast-seed-agents/
• Repository: https://github.com/Rohan5commit/storycast-seed-agents
• Successful GitHub render: https://github.com/Rohan5commit/storycast-seed-agents/actions/runs/24514194069
• Architecture: https://github.com/Rohan5commit/storycast-seed-agents/blob/main/docs/architecture.md
• Demo runbook: https://github.com/Rohan5commit/storycast-seed-agents/blob/main/docs/demo-runbook.md
• Demo video page: https://rohan5commit.github.io/storycast-seed-agents/demo.html

Recommended Challenge Area

Track 01: AI Video Agents

This is an inference from the public BetaHacks page as of April 18, 2026. StoryCast also overlaps with Content Automation, but AI Video Agents is the strongest fit.

One-Line Pitch

StoryCast is an autonomous multimodal agent that turns a single topic into a narrated one-minute short film using the BytePlus Seed generation stack.

Short Summary

StoryCast takes one plain-English idea like the death of a star and autonomously writes a scene-based narrative, generates storyboard images, synthesizes scene narration, animates each scene into video, and assembles the final film. The result is a complete explainer-style video from a single prompt.

Longer Project Description

StoryCast is built to show what a real video agent looks like when the orchestration itself is part of the product. A user provides one topic, and the system breaks it into a structured scene blueprint with narration, visual direction, motion cues, and tone metadata. Each scene becomes a storyboard frame, each frame becomes a video clip, each scene receives narration, and the pipeline merges everything into a polished 60-second film. The full process is visible, modular, and explainable, which makes it easy to demo and easy for judges to evaluate as true agentic execution rather than a single opaque generation call.

What Makes It Strong

• It creates a finished video artifact, not just intermediate assets.
• It makes the agent steps visible: planning, keyframing, narration, animation, and assembly.
• It is easy to understand in a two-minute demo because one input becomes one finished output.
• It is public and reproducible in GitHub, including a successful workflow render.

Technical Stack

• Seed 2.0: scene planning and narrative structure
• Seedream 5.0: keyframe image generation
• Seedance 2.0: image-to-video clip generation
• Seed Speech or temporary fallback TTS: scene narration
• ffmpeg: clip muxing and final concatenation
• FastAPI: demo surface
• GitHub Actions: cloud render path inside the public repository

Suggested Demo Video Structure

• 0:00-0:20: Open with the finished StoryCast video output
• 0:20-1:05: Show the user prompt, generation flow, and final result inside the app or GitHub workflow
• 1:05-1:35: Walk through the architecture and explain the model roles
• 1:35-2:00: Explain the future vision for education, science communication, and brand storytelling

Current Proof Points

• Public repo with documentation and CI
• Successful end-to-end GitHub render run
• Submission landing page
• Copy-paste demo runbook

Honest Compliance Note

The current successful public render used an ElevenLabs fallback for TTS because BytePlus Speech credentials were not available at run time. For strict BytePlus-only final compliance, switch the narration layer back to Seed Speech and rerun the pipeline with real BYTEPLUS_TTS_APP_ID, BYTEPLUS_TTS_TOKEN, and BYTEPLUS_TTS_CLUSTER values.

This site is open source. Improve this page.