Geminions

Inspiration

Design teams waste time on repetitive resizing, brand policing, and endless feedback cycles. We wanted a system that can generate assets fast without drifting off-brand, then critique and fix itself automatically.

What it does

Ingests brand guidelines + past assets into a Brand DNA Vault (colors, typography, tone) to lock in consistency. Generates high-fidelity visuals from a brief. Runs a Feedback Loop that critiques outputs for issues (wrong colors, illegible text, hallucinations) and triggers automatic edits. Adapts layouts across formats (banner → Story/Reel, carousel, etc.) while keeping text/logo inside safe zones. Enforces Narrative Continuity so characters/products remain consistent across a sequence. Extends static designs into lightweight motion (GIF/cinemagraph-style). Supports Google OAuth login with secure user accounts stored in Supabase.

How we built it

Frontend (React.js + Vite): UI for login, uploading brand guidelines/assets, entering briefs, previewing outputs, and exporting variants.

Backend (Node.js + Express.js): REST endpoints for brand ingestion, generation requests, critique/repair runs, and export orchestration, Handles session logic (3-hour login window) and route protection.

Auth + Database (Supabase + Google OAuth): Google OAuth for sign-in/sign-out, Supabase stores user profiles, brand vault metadata, briefs, generation history, and exported assets

Challenges we ran into

Brand fidelity: Keeping every output within strict brand rules while still allowing variety. Text legibility: Ensuring readable typography across crops and small placements. Cross-format adaptation: Reflowing layouts intelligently instead of naive resizing. Consistency across frames: Preventing character/product “drift” in multi-image sequences. Pipeline reliability: Orchestrating multiple steps (generate → critique → edit → export) without breaking the user experience.

Accomplishments that we're proud of

Built an end-to-end flow: brief → on-brand generation → critique → auto-fix → multi-format outputs. Implemented Google OAuth + Supabase user accounts for secure access. Created a scalable web service layer in Node/Express to orchestrate generation and refinement steps. Designed a React UI that makes a complex workflow feel simple.

What we learned

The best creative AI systems are agentic: generate, evaluate, repair, repeat. “Brand safety” needs both hard constraints (hex codes, typography rules) and soft constraints (tone/vibe). Adaptation is a layout problem, not a resizing problem—hierarchy and safe zones matter. Storing structured metadata in the vault makes consistency and iteration dramatically easier.

What's next for Geminions

More platform templates and safer auto-layout rules (Reels/TikTok, LinkedIn, email, display ads). Stronger Brand DNA ingestion (auto-detect logos, font matching, stricter token enforcement). Collaboration features: shared vaults, approvals, versioning, and audit trails. Better motion tools: timeline controls, transitions, and loop-aware animations. Analytics feedback: learn from what performs best and suggest improved variants automatically.