🖼️ DrawBench Glance More

This project performs inference over the DrawBench prompt benchmark using recent various image generation models. It enables automated image generation, logging, and grid visualization of outputs by prompt category.

---

🚀 Features

• 🔢 Inference using multiple SANA models (1.6B, 0.6B, 4.8B, etc.)
• 📊 Prompt categorization and controlled sampling
• 🧱 Grid visualization per category
• ✅ Supports all DrawBench categories or filtered categories
• 🔁 Reproducible via manual seed setting
• 🧩 Modular and extensible pipeline
• ⚡ Powered by uv for ultra-fast Python dependency management

---

🛠️ Quickstart

1. Clone the repository

git clone https://github.com/Longcat2957/DrawBench-Glance-More
cd DrawBench-Glance-More

2. Sync dependencies with uv

uv sync

This will automatically install all required packages listed in pyproject.toml.

---

📦 Directory Structure

.
├── prompts/
│   └── DrawBench.csv               # Prompt file grouped by categories
├── outputs/
│   └── {model_name}/{category}/   # Generated images
│   └── {model_name}/{category}_grid.png
├── models/
│   └── sana.py                     # SANA model loading logic
├── utils/
│   ├── logger.py                   # Logging configuration
│   ├── misc.py                     # Utility: device, dtype
│   └── grid.py                     # Grid image generation
└── main.py                         # Entry-point for prompt-based inference

---

🧪 Example Usage

🔹 Basic usage

uv run python main.py 

🔹 Run all categories

uv run python main.py --all-categories --num 100

---

🧠 Supported Models

Update --repo-id to try different preconfigured SANA models:

Model Repo ID	Inference Steps	Guidance Scale
Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers	2	—
Efficient-Large-Model/Sana_Sprint_0.6B_1024px_diffusers	2	—
Efficient-Large-Model/SANA1.5_4.8B_1024px_diffusers	20	4.5
Efficient-Large-Model/SANA1.5_1.6B_1024px_diffusers	20	4.5

---

🖼️ Output Example

Each category generates:

• Individual images named with sanitized prompt text and seed
• A grid image with all prompts from the category (unless --no_grid is set)

📁 Example:

outputs/
└── Sana_Sprint_1.6B/
    ├── Colors/
    │   ├── red_frog_on_leaf_seed42.png
    │   ├── ... 
    └── Colors_grid.png

---

📚 Prompt File Format

DrawBench.csv should follow:

Category,Prompt
Colors,A red frog on a green leaf
Colors,A blue elephant flying
Positional,A dog under the table
...

You may replace or expand this file for your own benchmarks.

---

🧑‍💻 Contributing

We welcome contributions to:

• Extend model support
• Add prompt sets (e.g., HRS, TIFA, etc.)
• Improve UI or grid rendering
• Optimize inference pipeline

Steps

1. Fork this repository
2. Create a feature branch
3. Open a pull request with clear description and sample output if applicable

---

📢 Acknowledgments

• The DrawBench benchmark
• SANA model family
• Contributors to the Diffusers ecosystem

---