An open-source text-to-image generator built with diffusion models and transformers

Imager is a modular, research-focused generative AI framework for creating high-quality images from text prompts. Built on the foundation of diffusion probabilistic models (DDPMs) and transformer architectures, it provides a transparent, customizable alternative to proprietary text-to-image systems.

Imager combines:

• Transformer encoder for text understanding (CLIP, T5, or custom language models)
• UNet-based diffusion decoder for iterative image generation through denoising
• Modular architecture enabling flexible experimentation and research

• Text Encoders: Support for pretrained and fine-tuned models
• Flexible Diffusion: Guided and unguided sampling strategies
• Custom Noise Schedules: Configurable denoising processes
• Multi-Resolution: Standard and super-resolution image generation
• Training Pipeline: From-scratch training and fine-tuning capabilities

• Implement core diffusion model architecture (UNet)
• Integrate text encoder (starting with CLIP)
• Create basic training pipeline
• Establish evaluation metrics and benchmarks

• Add support for multiple text encoders (T5, custom models)
• Implement advanced sampling strategies (DDIM, DPM-Solver)
• Create web interface for interactive generation
• Optimize inference speed and memory usage

• Support for custom dataset training
• Multi-modal conditioning (text + sketch, text + image)
• Advanced techniques (ControlNet, LoRA fine-tuning)
• Comprehensive documentation and tutorials

• Creative Applications: Art generation, design prototyping, storytelling
• Research: Multimodal learning, representation studies, generative modeling
• Data Augmentation: Synthetic dataset creation for ML pipelines
• Education: Understanding modern generative AI architectures

Unlike proprietary models, Imager prioritizes:

• Transparency: Full access to model architecture and training code
• Customizability: Easy modification for specific domains or research needs
• Reproducibility: Deterministic results and version-controlled experiments
• Community: Open collaboration for advancing generative AI research

Perfect for researchers, developers, and AI enthusiasts who want to understand, modify, and extend text-to-image generation systems.

Coming soon! The project is in early development.

This project welcomes contributions! Please see our contributing guidelines (coming soon).

This project is licensed under the terms specified in the LICENSE file.