wildminder/ComfyUI-DyPE

What it solves

ComfyUI-DyPE provides training-free methods to generate images at ultra-high resolutions (4K and beyond) using Diffusion Transformers (DiT). It prevents common high-resolution issues like spatial disorder (repeated or collapsed structures) and artifacts that occur when pushing pre-trained models far beyond their native training resolution.

How it works

The project implements several advanced patching techniques that can be inserted into a ComfyUI workflow after the model loader:

• Dynamic Position Extrapolation (DyPE): Adjusts positional encodings during the denoising process to handle low-frequency structures early and fine details later.
• Spatial Position Alignment (SPA): Fixes spatial disorder by bundling token positions and averaging attention outputs across multiple variants.
• Head-Adaptive attention Pruning (HAP): Accelerates inference by using block-sparse attention based on a pre-calibrated scope plan, reducing the computational load of high-resolution attention.
• SEGA: Uses content-aware spectral sharpening based on the latent's frequency spectrum.
• PixelRush & FreeScale: Provide cascade-based refinement and scale-fused attention for upscaling existing images or generating at higher resolutions.

Who it’s for

Digital artists and AI researchers using ComfyUI who want to generate high-fidelity 4K+ images using models like FLUX, Qwen Image, Z-Image, and Anima/Cosmos without needing to retrain the models.

Highlights

• Multi-Architecture Support: Works with FLUX, Nunchaku, Qwen Image, Krea-2, Z-Image, and Anima/Cosmos.
• Training-Free: No retraining or complex workflow changes required; simply patch the model via a single node.
• High-Resolution Capability: Enables generation at 4096×4096 and above.
• Inference Acceleration: HAP can provide up to 5.5x faster attention at 4K resolution.
• Zero Overhead: Many adjustments happen on-the-fly with negligible performance impact.