Diffusion Templates

A Unified Plugin Framework for Controllable Diffusion

Project Page | Paper | Code | Demo | Docs

Diffusion Templates decouples base-model inference from controllable capability injection. Template models map task-specific inputs to a standardized Template cache (KV-Cache, LoRA, etc.), which is then injected into the base diffusion pipeline — enabling reusable, composable plugins for controllable generation.

Resources

	Models	Datasets
🤖 ModelScope	KleinBase4B-Templates	ImagePulseV2
🤗 Hugging Face	KleinBase4B-Templates	ImagePulseV2

Model Zoo

11 Template models trained on FLUX.2-klein-base-4B, covering:

• Structural & Visual Control — ControlNet, Brightness, Color
• Editing & Attribute Alignment — Image Editing, Aesthetic, Age
• Enhancement & Reference — Super-Resolution, Content Reference, Sharpness
• Inpainting — Local Inpainting
• Easter Egg — Panda Meme 🐼

Datasets

ImagePulseV2: a large-scale open dataset collection (~1.2 TB) for training Diffusion Templates models, with 17 subsets covering text-to-image generation and diverse image editing tasks. All released under Apache License 2.0.

Citation

@article{duan2025diffusion,
  author    = {Duan, Zhongjie and Zhang, Hong and Chen, Yingda},
  title     = {Diffusion Templates: A Unified Plugin Framework for Controllable Diffusion},
  year      = {2025},
}

Acknowledgements

Website template from Nerfies (CC BY-SA 4.0).