HiDiffusion is a framework for improving the resolution and generation efficiency of pretrained diffusion models. Its core idea is to unlock higher-resolution, higher-quality, and faster image generation through minimal code changes without additional training.
Purpose Diffusion models have become the mainstream method for high-resolution image synthesis, but when directly using pretrained diffusion models to generate higher-resolution images, they encounter unreasonable object duplication and exponentially increasing generation time. HiDiffusion proposes solutions to these two major problems, enabling existing models to generate higher-resolution images at lower computational cost.
Main Capabilities
- Increased resolution: Supports expanding generation resolution up to 4096×4096, and demonstrates results on models such as SDXL, SDXL-Turbo, SD 2.1, and SD 1.5. Example resolutions include 1024×1024, 2048×2048, 2048×3072, and 4096×4096.
- Improved speed: Inference speed can reach 1.5-6 times that of previous methods, reducing the computational overhead of high-resolution generation.
- Solving object duplication: Dynamically adjusts feature map sizes through Resolution-Aware U-Net (RAU-Net) to alleviate the object duplication problem caused by duplication of deep features in the U-Net.
- Reducing self-attention redundancy: Optimizes window attention through Modified Shifted Window Multi-head Self-Attention (MSW-MSA) to reduce the computational burden caused by self-attention redundancy in the top layers of the U-Net.
- Task compatibility: Supports text-to-image generation, ControlNet tasks, and Inpainting tasks.
Usage The official description emphasizes that HiDiffusion can be integrated into various pretrained diffusion models by "adding just one line of code," thereby expanding image generation resolution and improving speed. Specific installation commands, code examples, and integration details are not yet provided. The official page displays a schematic diagram of the HiDiffusion framework, text-to-image generation tasks, image quality comparisons with other high-resolution image generation methods, efficiency comparisons with other acceleration methods, and examples of ControlNet and Inpainting tasks.
Applicable Scenarios Suitable for research or application scenarios that hope to use existing diffusion models (such as SDXL, SDXL-Turbo, SD 2.1, and SD 1.5) to generate high-resolution images while reducing generation time and avoiding object duplication. Can covertasks such as text-to-image generation, ControlNet control, and image inpainting (Inpainting).
Limitations The current materials do not provide specific installation steps, code examples, dependency environments, supported platforms, charging methods, or pricing information; these contents are not yet provided. Users need to refer to the official page for further technical details.
