diffgentor 0.1.1

A unified visual generation data synthesis factory supporting multiple backends (diffusers, xDiT, OpenAI API)

Diffgentor

A unified visual generation data synthesis tool for batch image generation and editing, designed for GenArena evaluation and beyond.

Abstract

Diffgentor is an efficient pipeline for batch image generation using various image generation and editing models. It supports multiple backends including diffusers, OpenAI API, Google GenAI (Gemini), and third-party models like Step1X-Edit, BAGEL, and Emu3.5.

Key features:

• Multiple Backends: diffusers, xDiT (multi-GPU), OpenAI, Google GenAI, and third-party models
• Batch Processing: Efficient batch inference with multi-process/multi-thread support
• GenArena Integration: Generate model outputs for GenArena pairwise evaluation
• Optimization Suite: VAE slicing/tiling, torch.compile, attention backends, and more

Quick Start

Installation

Option 1: pip install

# Core installation (diffusers, OpenAI, Google GenAI backends)
pip install diffgentor

# Install with all optional backends
pip install "diffgentor[all]"

GPU users: PyPI's default torch package is CPU-only. To use CUDA-enabled PyTorch, add the PyTorch index:

pip install diffgentor --extra-index-url https://download.pytorch.org/whl/cu126

flash-attn: The flash-attn optional dependency requires CUDA compilation. It is recommended to install a pre-built wheel manually:

pip install flash-attn --no-build-isolation

Or download a pre-built wheel from the flash-attention releases.

Option 2: From source (for development)

git clone https://github.com/ruihanglix/diffgentor.git
cd diffgentor
pip install -e ".[all]"

Download GenArena Dataset

hf download rhli/genarena --repo-type dataset --local-dir ./data

Generate Images for MultiRef Subset

Example using FLUX.2 [klein] 4B model:

diffgentor edit --backend diffusers \
    --model_name black-forest-labs/FLUX.2-klein-4B \
    --input ./data/multiref/ \
    --output_dir ./output/multiref/FLUX2-klein-4B/

Supported Backends

Backend	Type	Description
diffusers	T2I / Editing	HuggingFace diffusers with auto pipeline detection
xdit	T2I	Multi-GPU inference with xDiT parallelism
openai	T2I / Editing	OpenAI API (GPT-Image, DALL-E)
google_genai	T2I / Editing	Google GenAI (Gemini native image models)
step1x	Editing	Step1X-Edit model
bagel	Editing	ByteDance BAGEL model
emu35	Editing	BAAI Emu3.5 model
dreamomni2	Editing	DreamOmni2 (FLUX.1-Kontext + Qwen2.5-VL)
flux_kontext_official	Editing	BFL official Flux Kontext
hunyuan_image_3	Editing	Tencent HunyuanImage-3.0-Instruct

Documentation

Document	Description
Image Editing Guide	Comprehensive guide for image editing
Text-to-Image Guide	Text-to-image generation guide
Optimization Guide	Memory and speed optimization
Prompt Enhancement	LLM-based prompt enhancement
Environment Variables	Configuration via environment variables

Backend-Specific Guides

• Diffusers Models - Qwen, FLUX, LongCat
• Step1X-Edit - Step1X-Edit v1.0/v1.1
• BAGEL - ByteDance BAGEL
• Emu3.5 - BAAI Emu3.5
• DreamOmni2 - DreamOmni2
• Flux Kontext - BFL official
• HunyuanImage-3.0 - Tencent HunyuanImage
• OpenAI - GPT-Image API
• Google GenAI - Gemini

Environment Variables

Model-specific parameters are configured via DG_* environment variables:

# Step1X-Edit
DG_STEP1X_VERSION=v1.1
DG_STEP1X_SIZE_LEVEL=512

# BAGEL
DG_BAGEL_CFG_TEXT_SCALE=3.0
DG_BAGEL_CFG_IMG_SCALE=1.5

# API backends
OPENAI_API_KEY=your_key
GEMINI_API_KEY=your_key

See Environment Variables for the complete list.

Citation

@misc{li2026genarenaachievehumanalignedevaluation,
      title={GenArena: How Can We Achieve Human-Aligned Evaluation for Visual Generation Tasks?},
      author={Ruihang Li and Leigang Qu and Jingxu Zhang and Dongnan Gui and Mengde Xu and Xiaosong Zhang and Han Hu and Wenjie Wang and Jiaqi Wang},
      year={2026},
      eprint={2602.06013},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.06013},
}

License

Apache-2.0