Ad Generation Pipeline

A multi-modal ad generation pipeline. It pairs a local LLM (Ollama) for ad copy with a cloud image model (Hugging Face Inference API), wraps it in a Streamlit UI, and adds a ChromaDB memory loop that both remembers ads you liked and learns from how you edit drafts.

The default brand profile (src/brand.py) is pre-tuned to the brand's voice and audience, so out of the box it writes copy for products like video KYC, document verification, liveness detection, and background checks. Edit the brand live in the sidebar to retune.

Architecture

What it does

• Campaign generation — one product idea produces several distinct variants (benefit-led, problem-led, social-proof, urgency) for A/B testing. Choose 1–5 in the sidebar.
• Inline editing — tweak any variant's headline and body directly in the UI; no regeneration needed.
• Learns from your edits — "Teach from my edits" stores each draft→edit pair. Future campaigns pull the most semantically similar past edits in as few-shot examples, so copy drifts toward your preferred style over time.
• Liked-ads recall — "Save winner" stores an ad; as you type a new idea, past ads with a similar vibe surface automatically (semantic search, not keywords).
• Image style presets — render the visual for any variant in Photorealistic, Comic book, 3D render, Flat vector, or Watercolor styles (STYLE_PRESETS in src/prompts.py).

ad-generation-pipeline/
├── .env                 # Your secret HF token (NEVER commit). Create it yourself (see Quick start).
├── .gitignore           # Ignores .env, venv, chroma_db/, generated images, caches.
├── mental-model.svg     # Architecture diagram (shown above).
├── requirements.txt     # Python dependencies.
├── config.py            # Loads settings + secrets from .env.
├── app.py               # The Streamlit UI — run this.
└── src/
    ├── __init__.py
    ├── brand.py             # brand profile: voice, audience, values.
    ├── prompts.py           # LangChain prompt templates + image STYLE_PRESETS.
    ├── text_generator.py    # Ollama → campaign variants + image prompt.
    ├── image_generator.py   # Hugging Face → the actual image.
    └── feedback.py          # ChromaDB: liked-ads recall + learns-from-edits loop.

Quick start

python -m venv venv
source venv/bin/activate          # Windows: venv\Scripts\Activate.ps1
pip install -r requirements.txt

# Local LLM
# Install Ollama from https://ollama.com, then:
ollama pull llama3

# Secrets — create .env in the project root with your Hugging Face Read token:
#   HF_TOKEN=hf_your_token_here

Test each piece (in order)

python -m src.text_generator      # prints ad copy + a 3-variant campaign → LLM works
python -m src.image_generator     # saves generated_ad.png → images work
streamlit run app.py              # opens the full UI

Notes

• Ollama runs locally at http://localhost:11434 — no API key, no per-request cost. If Python says "connection refused", run ollama serve and confirm the model name in config.py matches ollama list exactly (case-sensitive).
• Hugging Face is the only cloud piece. A free Read token works for prototyping. A 401 means the token is missing/wrong; never commit .env.
• Image model fallbacks — if stabilityai/stable-diffusion-xl-base-1.0 404s, switch IMAGE_MODEL in config.py to black-forest-labs/FLUX.1-schnell or stabilityai/stable-diffusion-2-1. Check a model's page for the "Inference Providers" widget to confirm it's callable.
• ChromaDB stores liked ads and draft→edit pairs in chroma_db/ (gitignored) and finds similar ones by meaning, not keywords. Both the recall and the learning loop are cold-start safe — with nothing saved yet, they simply contribute nothing.