F L U X.1 Turbo Alpha

Introduction

FLUX.1-Turbo-Alpha is an advanced text-to-image model developed by Alimama-Creative, based on the FLUX.1-dev framework. It incorporates a multi-head discriminator to enhance the quality of image generation and is suitable for various applications including text-to-image transformation, inpainting, and controlnet tasks.

Architecture

The model uses a distilled Lora approach with an 8-step process, enhancing the original FLUX.1-dev model. It employs a multi-head discriminator architecture to improve distillation quality. The model is specifically designed for tasks such as text-to-image generation and inpainting controlnets.

Training

FLUX.1-Turbo-Alpha is trained on a dataset of 1 million images sourced from both open and internal channels. These images are selected based on aesthetic scores of 6.3+ and resolutions above 800x800 pixels. The model uses adversarial training with the original FLUX.1-dev transformer as the discriminator backbone and adds multi heads to each transformer layer. Key training parameters include:

• Mixed precision: bf16
• Learning rate: 2e-5
• Batch size: 64
• Image size: 1024x1024
• Guidance scale: 3.5
• Time shift: 3

Guide: Running Locally

To run FLUX.1-Turbo-Alpha locally, follow these steps:

1. Install Dependencies: Ensure you have the diffusers library installed.
2. Load Model:

   import torch
   from diffusers.pipelines import FluxPipeline
   
   model_id = "black-forest-labs/FLUX.1-dev"
   adapter_id = "alimama-creative/FLUX.1-Turbo-Alpha"
   
   pipe = FluxPipeline.from_pretrained(
     model_id,
     torch_dtype=torch.bfloat16
   )
   pipe.to("cuda")
   
   pipe.load_lora_weights(adapter_id)
   pipe.fuse_lora()
   

3. Generate Image: Use a prompt to create images.

   prompt = "A DSLR photo of a shiny VW van with a cityscape painted on it..."
   image = pipe(
               prompt=prompt,
               guidance_scale=3.5,
               height=1024,
               width=1024,
               num_inference_steps=8
              ).images[0]
   

4. Suggested Environment: For optimal performance, consider using cloud GPUs like AWS EC2, Google Cloud GPUs, or Azure NV-series.

License

FLUX.1-Turbo-Alpha is distributed under the flux-1-dev-non-commercial-license. For more details, please refer to the license link.