Agentic Brand Guardian

Inspiration

Brands struggle to maintain visual consistency, compliance, and creativity across campaigns. Marketing teams spend hours validating content, checking design guidelines, and generating visuals that meet brand standards. We wanted to build a platform that automates brand validation while enabling creative, high-quality image generation — combining AI agents, LLMs, and FIBO’s precise visual control.

What it does

Agentic Brand Guardian is an AI-powered workflow that:

Validates brand descriptions against a set of rules and guidelines.

Enhances textual descriptions to make them more appealing and professional.

Generates high-quality images using Bria FIBO with control over camera angles, lighting, FOV, and color palette.

Accepts JSON-based inputs to automate production-ready visual workflows.

Provides a Streamlit interface for easy interaction by marketing and creative teams.

It effectively combines automation, AI creativity, and brand compliance into one intelligent tool.

How we built it

LLM Integration: Used Hugging Face models to improve brand description quality.

Agentic Validation: A “Brand Agent” validates descriptions against rules in brand_schema.json.

FIBO Generator: Integrated Bria FIBO (or CPU-friendly Stable Diffusion fallback) for JSON-native image generation.

Streamlit UI: Built a responsive interface for users to input data, adjust image parameters, and generate results.

JSON Schema: Structured input (image_schema.json) to enable repeatable, automated workflows.

Colab Deployment: Runs in Google Colab with GPU support using pyngrok to expose Streamlit.

Challenges we ran into

Model size & memory: Bria FIBO models are large; running on CPU caused MemoryError. Solved using GPU in Colab or lightweight SD models.

Token and gated repo access: Hugging Face HF_TOKEN management required careful environment setup.

Integration of LLM + image pipeline: Ensuring smooth data flow from text validation/enhancement to JSON-controlled image generation.

Latency on GPU/CPU: Adjusted inference steps to balance quality vs runtime.

Accomplishments that we're proud of

What we learned

Developed a fully agentic workflow from brand validation to professional image generation.

Achieved JSON-native control over all aspects of image generation.

Built a Colab-ready, cloud-accessible Streamlit demo for easy testing and competition submission.

Created an end-to-end pipeline that is scalable, reproducible, and production-ready.

Built With

• llms
• python