ComfyUI-DyPE

Why we included this project

Diffusion transformer models have a native resolution ceiling, and once you push past it the output tends to fall apart. This ComfyUI node pack patches FLUX, Qwen Image, Z-Image, Anima/Cosmos, and Krea-2 right after the loader, so you can generate at 2K, 4K, or higher without retraining or touching your existing workflow. The DyPE node handles the core work by extrapolating positional encodings during denoising, SEGA is a spectral-sharpening alternative, and the HRDiT pair splits the remaining problems: SPA fixes repeated or collapsed structures at high resolution, HAP accelerates sparse attention. Everything is training-free and drops in as a single node, so it plays nice with the samplers and optimization nodes you already use. If your output looks soft at native resolution, this is a cheap way to test whether the model itself is the limit.