| Property | Value |
|---|---|
| License | FLUX-1-dev-non-commercial-license |
| Base Model | FLUX.1-dev |
| Library | Diffusers |
| Training Precision | BF16 |
What is FLUX.1-Turbo-Alpha?
FLUX.1-Turbo-Alpha is an advanced 8-step distilled LoRA model developed by the AlimamaCreative Team. Built upon the FLUX.1-dev foundation, it represents a significant optimization in text-to-image generation, achieving high-quality outputs with remarkably fewer inference steps. The model utilizes a sophisticated multi-head discriminator architecture to maintain generation quality while significantly improving speed.
Implementation Details
The model was trained on a curated dataset of 1M images, filtered for aesthetic scores above 6.3 and resolutions exceeding 800 pixels. Key technical specifications include training with bf16 mixed precision, a learning rate of 2e-5, and a batch size of 64. The model operates optimally with a guidance scale of 3.5 and lora_scale of 1.
- Multi-head discriminator implementation for enhanced quality control
- Adversarial training methodology with fixed transformer backbone
- Time shift parameter of 3 during training
- Support for 1024x1024 resolution outputs
Core Capabilities
- Efficient text-to-image generation in just 8 steps
- Compatible with inpainting controlnet applications
- Supports high-resolution image generation
- Seamless integration with diffusers library
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its ability to generate high-quality images in just 8 inference steps, achieved through sophisticated distillation techniques and multi-head discriminator training.
Q: What are the recommended use cases?
The model excels in rapid text-to-image generation and is particularly well-suited for applications requiring quick turnaround times while maintaining quality. It's also effective for inpainting tasks when combined with appropriate controlnet models.