Diet AI — Flan-T5-XL Version

Better quality nutrition analysis using Google's Flan-T5-XL (3 billion parameters). No API keys. No accounts. 100% HuggingFace open-source.

Models used

Model	Size	Purpose	RAM needed
openai/clip-vit-base-patch32	~600MB	Food image recognition	~800MB
google/flan-t5-xl	~3GB	Nutrition analysis	~4GB
Open Food Facts	Free API	Real nutrition data	0

Total RAM needed: ~5GB

---

Why XL over Large?

	Flan-T5-Large	Flan-T5-XL
Parameters	770M	3B
Size	770MB	3GB
RAM needed	~1GB	~4GB
Streamlit Cloud	Free tier	Paid tier / local
Nutrition accuracy	Good	Much better
JSON instruction following	Decent	Excellent
Speed on CPU	~20s/query	~45s/query

---

Quick start

# 1. Make sure you have 6GB+ free RAM and 6GB+ free disk

# 2. Go into the project folder
cd diet_ai_xl

# 3. Create virtual environment
python -m venv venv
source venv/bin/activate      # Mac/Linux
venv\Scripts\activate         # Windows

# 4. Install packages
pip install -r requirements.txt

# 5. Run
streamlit run app.py
# First run downloads ~3.6GB total — takes 5-10 minutes once

---

Swap models easily

Change one line in diet_classifier.py:

# Free tier / fast:
MODEL_NAME = "google/flan-t5-large"    # 770MB, ~1GB RAM

# This version (better quality):
MODEL_NAME = "google/flan-t5-xl"       # 3GB, ~4GB RAM

# Maximum quality (GPU recommended):
MODEL_NAME = "google/flan-t5-xxl"      # 11GB, ~14GB RAM

---

GPU acceleration (optional but faster)

If you have an NVIDIA GPU, change one line in diet_classifier.py:

# Change:
device="cpu"

# To:
device=0    # uses your first GPU

Also install CUDA PyTorch:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

---

Troubleshooting

Problem	Fix
Killed or out of memory	Close other apps, need 5GB+ free RAM
Very slow (>2 min/query)	Normal on CPU — use GPU or switch to flan-t5-large
First run hangs at 0%	Slow internet — wait, model is 3GB
ModuleNotFoundError	pip install -r requirements.txt
Streamlit Cloud crashes	XL needs paid tier — use flan-t5-large for free tier