HiFlow: Training-free High-Resolution Image Generation with Flow-Aligned Guidance (2504.06232v2)

Abstract: Text-to-image (T2I) diffusion/flow models have drawn considerable attention recently due to their remarkable ability to deliver flexible visual creations. Still, high-resolution image synthesis presents formidable challenges due to the scarcity and complexity of high-resolution content. Recent approaches have investigated training-free strategies to enable high-resolution image synthesis with pre-trained models. However, these techniques often struggle with generating high-quality visuals and tend to exhibit artifacts or low-fidelity details, as they typically rely solely on the endpoint of the low-resolution sampling trajectory while neglecting intermediate states that are critical for preserving structure and synthesizing finer detail. To this end, we present HiFlow, a training-free and model-agnostic framework to unlock the resolution potential of pre-trained flow models. Specifically, HiFlow establishes a virtual reference flow within the high-resolution space that effectively captures the characteristics of low-resolution flow information, offering guidance for high-resolution generation through three key aspects: initialization alignment for low-frequency consistency, direction alignment for structure preservation, and acceleration alignment for detail fidelity. By leveraging such flow-aligned guidance, HiFlow substantially elevates the quality of high-resolution image synthesis of T2I models and demonstrates versatility across their personalized variants. Extensive experiments validate HiFlow's capability in achieving superior high-resolution image quality over state-of-the-art methods.

• The paper introduces HiFlow, a training-free framework that enables pre-trained flow models to generate high-resolution images without requiring additional training.
• HiFlow utilizes a virtual reference flow and aligns initialization, denoising direction, and acceleration during sampling to preserve structural features and details.
• Experimental results demonstrate that HiFlow achieves state-of-the-art performance on metrics like FID, IS, and CLIP, and is model-agnostic, compatible with various T2I systems.

Overview of HiFlow: A Training-Free Approach to High-Resolution Image Generation

The paper introduces HiFlow, a novel framework designed to enhance the performance of pre-trained flow models in generating high-resolution images without the need for additional training. This approach addresses a significant challenge in the domain of text-to-image (T2I) generation: the deterioration of image quality when models are tasked with synthesizing images at resolutions higher than they were originally trained for.

Key Contributions and Methodology

HiFlow operates by establishing a virtual reference flow within the high-resolution space, utilizing information from the low-resolution sampling flow. This reference provides essential guidance that improves high-resolution image generation across three main dimensions:

1. Initialization Alignment: HiFlow initiates the high-resolution sampling with a virtual noisy image that correlates with the upsampled low-resolution image. This ensures that the generated high-resolution outputs maintain consistency with their low-resolution counterparts in terms of low-frequency content.
2. Direction Alignment: By aligning the denoising direction with that of the reference flow, the method effectively preserves structural features throughout the image generation process. This alignment is achieved by adjusting the low-frequency components of the generated image using Fourier-based techniques to match those of the reference image.
3. Acceleration Alignment: HiFlow aligns the acceleration of the flow (the second derivative of the image generation trajectory) to ensure fidelity in detail reproduction. This aspect of the method prevents the introduction of unrealistic patterns and ensures that the timing and sequence of detail generation are consistent with the model's inherent preferences.

Experimental Validation

The paper presents extensive experimental results, positioning HiFlow as superior to current state-of-the-art methods in high-resolution image synthesis. The evaluation metrics include FID, IS, and CLIP scores, indicating HiFlow's performance in maintaining image quality and ensuring alignment with textual prompts. Numerical results confirm significant improvements over alternative techniques, including training-free and training-based approaches for generating high-resolution images.

Practical and Theoretical Implications

HiFlow’s model-agnostic nature allows its application across various architectures without requiring model-specific adaptations, enabling broad utility for different T2I systems. This versatility extends to customized T2I models using LoRA or ControlNet and even quantized models, highlighting the framework's adaptability and practical significance in scenarios where computational resources are constrained.

From a theoretical perspective, HiFlow challenges the conventional necessity of fine-tuning large-scale models with high-resolution datasets, offering a paradigm shift through its innovative use of trajectory-aligned guidance. The method underscores a promising direction for future research in AI, particularly in the development of efficient, scalable models capable of extending their utility beyond their original training capabilities.

Conclusion

HiFlow stands out as a method that successfully extends the capabilities of diffusion and flow models in high-resolution image generation, offering a cost-effective, training-free solution. Its ability to preserve detail and structural integrity while adapting pre-trained models for new high-resolution tasks marks a noteworthy advancement in the field, with far-reaching implications for both academic research and practical applications in AI-driven image synthesis. Future developments may further refine this approach by integrating it with newer architectures and exploring additional forms of trajectory guidance.