Inspiration
The traditional pre-visualization (pre-vis) process in filmmaking is time-consuming and expensive, often requiring skilled artists to manually create storyboards that can cost studios tens of thousands per sequence. With the rise of AI image generation, tools like Midjourney offer creativity but lack the deterministic control needed for professional workflows—outputs vary unpredictably, breaking consistency across shots. Bria FIBO's groundbreaking JSON-native architecture changes this by providing precise, structured control over professional parameters (camera angles, FOV, lighting, color palettes) via its LLM translator and disentangled generation. Inspired by FIBO's potential to revolutionize enterprise visual AI, we envisioned SceneForge as an agentic tool that automates pre-vis for directors and VFX teams, turning high-level briefs into consistent, production-ready HDR storyboards.
What it does
SceneForge is a web-based agentic pipeline that transforms natural-language scene descriptions (e.g., "A dramatic cyberpunk chase in pouring rain at night") into full pre-vis storyboards: sequences of 8-20 high-fidelity images with perfect multi-shot consistency. An "orchestra" of AI agents breaks down the script, translates it into FIBO's structured JSON (locking elements like character appearance, volumetric lighting, and 16-bit HDR palettes), generates frames, and allows iterative refinements via targeted JSON edits. Users get an interactive timeline for reordering/annotating shots, exports in EXR format for Nuke integration, and pro-grade outputs safe for commercial use thanks to FIBO's licensed training data.
How we built it
We built SceneForge with a modern, production-ready stack:
- Frontend: React (Vite + TypeScript) with Tailwind CSS for a stunning dark-mode UI, Framer Motion for smooth animations, ShadCN components for sliders/modals, and a custom interactive timeline for storyboard viewing.
- Backend: Python FastAPI handling the heavy lifting—integrating Bria FIBO via its API (for cloud-based generation without local GPU needs), LangChain for the agentic workflow (script breakdown, JSON structuring, consistency enforcement, refinement agents), and Bria's LLM translator for natural-to-JSON conversion.
- Key Integrations: FIBO API for deterministic generations with pro parameters; OpenEXR exports for 16-bit HDR; mock modes for demo reliability. The split allowed a polished, responsive UX while leveraging Python's strength for AI orchestration—all in a single repo, deployed locally or via cloud for the submission demo.
Challenges we ran into
Time constraints in a hackathon sprint made balancing ambition with feasibility tough—initially scoping VR previews but pivoting to core agentic flow. Integrating FIBO's API required careful handling of rate limits and JSON schema validation to avoid generation failures. Ensuring cross-frame consistency (e.g., identical character poses/lighting) demanded clever agent prompting and parameter locking, with early iterations showing minor drifts we fixed via a dedicated "consistency agent." Heavy image outputs slowed the UI; we optimized with lazy loading, skeletons, and progress tracking for agent steps.
Accomplishments that we're proud of
- Creating a truly agentic, end-to-end pipeline that showcases FIBO's JSON-native strengths in a scalable, production-relevant way—generating consistent HDR storyboards in minutes, something manual tools take days for.
- A beautiful, intuitive React frontend with live JSON editing, param sliders, and timeline interactions that feels like a pro tool (not a hack prototype).
- Demonstrating real-world impact: Outputs are Nuke-ready EXR sequences with 16-bit depth, leveraging FIBO's commercial indemnity for enterprise safety.
- Completing a polished demo video showing iconic scenes (e.g., noir chase) turned into pre-vis reels, highlighting deterministic control over randomness.
What we learned
We deepened our understanding of agentic AI workflows—chaining LLMs with structured generation unlocks reliability far beyond prompt engineering. FIBO's disentangled parameters make iterative refinement feel magical, reinforcing how JSON-native control is the future for professional creative AI. Building split-stack apps (React + FastAPI) taught us the power of separation for performance and UX polish. Finally, targeting real production pain points (like film pre-vis costs) amplifies a project's judging impact in AI hackathons.
What's next for SceneForge
Post-hackathon, we'll add multi-user collaboration (shareable sessions for director feedback), deeper integrations (direct Nuke/Blender exports, ComfyUI nodes), and fine-tuning on user assets for brand-consistent pre-vis. Exploring image-to-JSON inspiration (using FIBO's VLM for reference uploads) and video extension via frame interpolation. Ultimately, aim for open-source release with Bria partnership, evolving into a full pre-vis platform for indie filmmakers and studios alike.
Log in or sign up for Devpost to join the conversation.