Campus-AI — Architecture

01

Data Pipeline

Collect → Filter → Caption → Split

🕷️

Pinterest Scraper

Selenium headless browser with automatic scrolling & perceptual hash deduplication

57 subcategories × 1,900

🔬

Quality Filter

GPU-accelerated Laplacian sharpness, resolution, aspect ratio & color diversity checks

~68% pass rate

📝

Florence-2 Captioner

Microsoft Florence-2-large generates detailed captions in bf16 with torch.compile

SM120 optimized

✂️

Dataset Splitter

Stratified splitting by category into training, validation & test sets

~55K train images

02

Training Pipeline

Fine-tune Flux.1-dev with LoRA adapters

⚡ Prodigy Optimizer

Self-adapting LR = 1.0
No manual LR tuning needed

📉 Min-SNR-γ Loss

γ = 5.0 — balanced learning
across all noise levels

🔄 Cosine Warm Restarts

3 cycles over 4 epochs
escapes local minima

🧠

Flux.1-dev

12 billion parameter
transformer diffusion model

+ LoRA Adapter (Rank 16, α=16)

40M

Trainable Params

bf16

Precision

4

Eff. Batch Size

~55K

Optimizer Steps

🛡️ Anti-Overfitting

Caption dropout 10%
LoRA dropout 8%
L2 weight decay 0.01

⚙️ LoRA+ (ICML '24)

B matrix gets 16× higher LR
Free +2% accuracy boost

🖥️ SM120 Blackwell

TF32 tensor cores
torch.compile max-autotune

03

Inference & Deployment

Prompt → Generate → Upscale → Deliver

👤

User Input

Event description, type, visual style & resolution preset

🦙

Groq Llama 3.3 70B

Enhances plain text into detailed Flux-optimized prompts

~200ms API

Flux.1-dev + LoRA Inference Engine

✍️

Text → Poster

From description only

🖼️

Reference Style

IP-Adapter transfer

🔄

Image → Image

Transform existing art

🎭

Inpainting

Edit specific regions

🔎

Real-ESRGAN 2×

AI upscaling for crisp HD output at any size

🎨

Generated Poster

1024×1024 to 1152×768
Multiple variants supported

🖥️ Local — RTX 5070 Ti (12GB VRAM)

☁️ Cloud — HF Spaces + ZeroGPU