LavenderFlow 5.6B
| Property | Value |
|---|---|
| Parameter Count | 5.6B |
| License | MIT |
| Resolution | 768x768 |
| Training Duration | ~3 weeks |
| Paper Reference | SSCD embeddings paper |
What is lavenderflow-5.6B?
LavenderFlow 5.6B is an open-source text-to-image generation model developed by a single graduate student in just three weeks. It represents a significant achievement in democratizing AI development, proving that foundation models aren't exclusively the domain of large companies. The model incorporates advanced features like Latent diffusion, MMDiT, muP, and CFM, while utilizing FSDP for efficient training.
Implementation Details
The model was trained using a single 8xH100 node configuration, processing approximately 550,000 steps with a batch size of 128. It leverages the T5-large language model for text encoding and implements DeepSpeed with Zero algorithm stage-2 for efficient training, achieving 45-60% MFU (Model FLOPs Utilization).
- Trained on CapFusion Dataset and ye-pop dataset
- Implements SSCD embeddings and FAISS for deduplication
- Utilizes SDXL-VAE embeddings for 256x256 image processing
- Supports high-resolution 768x768 image generation
Core Capabilities
- Text-to-image generation at 768x768 resolution
- Efficient processing using T5 integration
- Optimized training through muTransfer and FSDP
- Balanced performance despite limited training resources
Frequently Asked Questions
Q: What makes this model unique?
This model demonstrates that complex AI systems can be developed by individual researchers with limited resources, challenging the notion that foundation models require large teams and extensive infrastructure.
Q: What are the recommended use cases?
While the model is currently described as "severely undertrained," it's suitable for experimental text-to-image generation tasks and research purposes. It's important to note that it's not intended to compete with state-of-the-art models like SD3, Midjourney, or DALLE-3.
