Summary

This ComfyUI custom node integrates DyPE (Dynamic Position Extrapolation) to enable FLUX-based diffusion models to generate images at resolutions far exceeding their training data, such as 4K and beyond. It addresses common artifacts and coherence issues encountered during ultra-high-resolution generation, offering a training-free enhancement with zero inference overhead for ComfyUI users.

How It Works

DyPE is a novel, training-free method that dynamically adjusts a diffusion model's positional encodings at each sampling step. By aligning the frequency spectrum of these encodings with the current stage of the generative process—prioritizing low frequencies early and high frequencies later—DyPE prevents the structural degradation and repeating artifacts typically seen when pushing models to extreme resolutions. This single-node integration seamlessly patches FLUX models, allowing them to generate high-resolution outputs without additional computational cost.

Quick Start & Requirements

Installation is straightforward via the ComfyUI Manager by searching for "ComfyUI-DyPE" or manually by cloning the repository into the ComfyUI/custom_nodes/ directory. No additional dependencies beyond ComfyUI and a FLUX-based model are required.

Highlighted Details

• Enables true high-resolution generation (4096x4096+) with maintained global coherence and detail.
• Features a simple, single-node "plug-and-play" integration into existing ComfyUI workflows.
• Offers full compatibility with standard ComfyUI samplers, schedulers, and optimization nodes.
• Provides fine-grained control over DyPE hyperparameters, including dype_exponent for tuning extrapolation strength.
• Achieves zero inference overhead, with negligible performance impact during generation.

Maintenance & Community

Community discussions and support can be found via the 'TokenDiff AI News' and 'TokenDiff Community Hub'.

Licensing & Compatibility

The underlying DyPE method is patent pending. Commercial use or licensing inquiries for the core technology should be directed to the original authors. This node is exclusively compatible with FLUX-based model architectures and will not function with standard U-Net models like SD 1.5 or SDXL.

Limitations & Caveats

The node is strictly limited to FLUX model architectures. A known issue exists with the node's width/height parameters, which are reportedly buggy; users should keep these values below 1024x1024, as they do not impact the final output resolution. Optimal parameter tuning, particularly dype_exponent, requires experimentation based on the target resolution.