1roOt/qr-code-illusions-30k

🔮 QR Code Illusions 30k

A high-fidelity, verified paired dataset of 30,000 artistic QR Code illusions, complete with multi-decoder verification and quantitative robustness scoring for training ControlNet and Flow-Matching models.

[image]

🌟 Overview

The QR Code Illusions 30k dataset is designed to train conditional diffusion models (ControlNet, ControlNet-LLLite, LoRA, and Flow-Matching MMDiT adapters) capable of embedding scannable QR codes into aesthetic artwork.

Unlike synthetic or naive QR blends, each sample in this dataset represents a fine-tuned optical illusion where:

1. 1.The QR code is seamlessly embedded into architectural structures, natural foliage, lighting highlights, and artistic textures.
2. 2.Visual aesthetics are preserved without harsh black/white square artifacts.
3. 3.Every image is rigorously validated across multiple real-world QR decoders (pyzbar, OpenCV WeChat QR) and tested against physical distortions (rotation, perspective shift, blur, compression).

🎨 Featured Samples

Here are four representative examples selected directly from the dataset:

📊 Dataset Structure & Statistics

Splits (Zero Matrix Leakage)

To prevent model memorization, the dataset was partitioned using QR matrix hash bucketing ensuring no identical QR layout appears across train and evaluation sets:

• —`train`: 23,988 pairs
• —`validation`: 3,021 pairs
• —`test`: 1,486 pairs
• —`hard_test`: 1,505 pairs (high-difficulty scanning scenarios)
• —Total: 30,000 paired samples

Quality Tiers

• —Class A (10,564 pairs): Highly scannable under standard and adverse conditions; excellent visual fidelity.
• —Class B (5,576 pairs): Reliable scan under standard conditions; enhanced artistic subtlety.
• —Class C (13,860 pairs): High artistic immersion and visual blending (useful for contrast and texture pre-training).
• —Note: Corrupt and unreadable Class D samples have been strictly removed.

📋 Schema & Metadata Fields

Each sample contains:

🚀 Quickstart & Usage

1. Load with Hugging Face datasets

from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("1roOt/qr-code-illusions-30k")

# Access train sample
sample = dataset["train"][0]
image = sample["image"]          # PIL.Image
condition = sample["condition"]  # PIL.Image
prompt = sample["prompt"]        # str

print(f"Loaded: {sample['id']} | Style: {sample['style']} | Class: {sample['quality_class']}")

2. Filter for High-Scannability Samples (Class A only)

# Filter for reliable scannability training
class_a_train = dataset["train"].filter(lambda x: x["quality_class"] == "A")
print(f"Number of Class A training samples: {len(class_a_train)}")

3. PyTorch DataLoader Example for ControlNet Training

from torch.utils.data import DataLoader
from torchvision import transforms

transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize([0.5], [0.5])
])

def collate_fn(batch):
    images = torch.stack([transform(b["image"]) for b in batch])
    conditions = torch.stack([transform(b["condition"]) for b in batch])
    prompts = [b["prompt"] for b in batch]
    return {"images": images, "conditions": conditions, "prompts": prompts}

loader = DataLoader(dataset["train"], batch_size=4, shuffle=True, collate_fn=collate_fn)

⚖️ License & Attribution

• —Generated Images: Base generations derived from Stable Diffusion architectures under the CreativeML OpenRAIL-M license.
• —QR Synthesis: Generated algorithmically via open-source tools under MIT / BSD licenses.
• —Dataset & Metadata: Released under OpenRAIL / Creative Commons Attribution 4.0 (CC BY 4.0).