SNOOPI: Setting a New Benchmark for One-Step Diffusion Models

Introducing Dynamic Guidance and Negative Prompt Integration for Superior Image Generation

• Enhanced Stability: SNOOPI introduces Proper Guidance – SwiftBrush (PG-SB) to stabilize training by dynamically adjusting the guidance scale across diverse model backbones.
• Innovative Control: The Negative-Away Steer Attention (NASA) method integrates negative prompt guidance, enabling suppression of unwanted elements in generated images.
• State-of-the-Art Performance: SNOOPI achieves a record-breaking HPSv2 score of 31.08, redefining benchmarks for one-step diffusion models.

The demand for efficient and high-quality text-to-image generation models has driven significant advancements in diffusion technologies. However, current one-step diffusion models face challenges such as instability across different backbones and the lack of support for negative prompt guidance. Enter SNOOPI, a groundbreaking framework that addresses these limitations with innovative training and inference strategies.

SNOOPI builds on the success of methods like SwiftBrushv2 (SBv2), which demonstrated the potential of one-step distillation but struggled with stability and practical usability. By introducing Proper Guidance – SwiftBrush (PG-SB)and the Negative-Away Steer Attention (NASA) method, SNOOPI achieves unparalleled performance and control, pushing the boundaries of what one-step diffusion models can accomplish.

Key Innovations in SNOOPI

1. Proper Guidance – SwiftBrush (PG-SB): Stabilizing Training

PG-SB dynamically adjusts the guidance scale during training, broadening the teacher model’s output distributions. This random-scale classifier-free guidance ensures a more robust Variational Score Distillation (VSD) loss, allowing SNOOPI to perform consistently across diverse model backbones without additional computational costs.

2. Negative-Away Steer Attention (NASA): Practical Control in Image Generation

One of SNOOPI’s most significant contributions is NASA, the first-ever method to integrate negative prompt guidance in one-step diffusion models. By leveraging cross-attention adjustments, NASA effectively suppresses unwanted features, providing creators with greater control over generated images.

Performance and Benchmarking

SNOOPI’s innovations translate into tangible improvements, setting new records in one-step diffusion performance:

• HPSv2 Score: Achieved a state-of-the-art score of 31.08, surpassing existing baselines.
• Versatility: Demonstrated superior performance across diverse backbones, addressing the instability issues observed in previous models like SBv2.
• Practical Utility: Enhanced usability through negative prompt guidance, addressing a critical gap in practical image generation.

A New Standard for One-Step Diffusion Models

SNOOPI represents a transformative step in one-step diffusion technologies, combining enhanced stability with unprecedented control. With its innovative PG-SB training method and NASA’s negative prompt integration, SNOOPI delivers a robust, versatile, and user-friendly framework for high-quality image generation.

As diffusion models continue to evolve, SNOOPI sets a new benchmark for efficiency and effectiveness, paving the way for further innovations in AI-driven creativity and visual intelligence.