SDXL Turbo

SDXL Turbo represents a significant advancement in real-time text-to-image generation, introducing a novel distillation technique called Adversarial Diffusion Distillation (ADD) that enables high-quality image synthesis in a single step. As detailed in the research paper, this breakthrough allows for dramatically faster generation compared to its predecessor SDXL 1.0, which required 50 steps for comparable results.

Technical Architecture and Innovation

The model's architecture leverages ADD, which combines adversarial training with score distillation from a pre-trained teacher model. This approach maintains high image fidelity while enabling rapid generation. The ADD framework consists of three key components:

• An ADD-student network (initialized with pre-trained weights)
• A discriminator network
• A frozen teacher diffusion model

The discriminator is conditioned on both text and image embeddings to enhance performance, utilizing Vision Transformers (ViTs) trained with the DINOv2 objective. This architecture allows for iterative refinement, with image quality improving through additional sampling steps when desired.

Two variants were developed during research:

• ADD-M: 860M parameters
• ADD-XL: 3.1B parameters (SDXL backbone)

Performance and Capabilities

SDXL Turbo demonstrates remarkable performance in both single-step and multi-step configurations. On an A100 GPU, it generates 512x512 images in approximately 207ms, with a single UNet forward evaluation taking just 67ms. Blind tests against state-of-the-art models revealed that SDXL Turbo outperforms:

• A 4-step LCM-XL configuration with just a single step
• A 50-step SDXL configuration with only 4 steps

The model excels at producing photorealistic images but has some limitations:

• Fixed resolution of 512x512
• Difficulty rendering legible text
• Potential issues with faces and people
• Lossy autoencoding artifacts

Implementation and Usage

SDXL Turbo can be implemented using the Diffusers library. The basic setup requires:

pip install diffusers transformers accelerate --upgrade

For text-to-image generation:

AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", 
    torch_dtype=torch.float16, 
    variant="fp16")

Optimal parameters include:

• num_inference_steps=1
• guidance_scale=0.0

For image-to-image generation, use AutoPipelineForImage2Image, ensuring num_inference_steps * strength >= 1.

The model is licensed under the sai-nc-community license for non-commercial and research purposes. Primary applications include:

• Research on generative models
• Real-time applications
• Artistic creation
• Design tools
• Educational applications

References

1. SDXL Turbo Research Paper - Technical details of the ADD technique
2. Model Weights and Code - Official model repository
3. Generative Models Repository - Implementation details
4. SDXL 1.0 Base Model - Original model that SDXL Turbo was derived from
5. License Information - Detailed licensing terms