wildminder/ComfyUI-DyPE

ComfyUI‑DyPE – High‑resolution diffusion for ComfyUI

What it is – A collection of custom nodes for the visual‑programming UI ComfyUI that let you push modern diffusion‑transformer models (Flux, Qwen‑Image, Z‑Image, Anima/Cosmos, Krea‑2, etc.) to ultra‑high resolutions (2 K, 4 K and beyond) without retraining the model.

How it works – The nodes implement training‑free patches that modify the model’s positional encodings or attention patterns at inference time:

• DyPE – Dynamic Position Extrapolation, adjusts RoPE on the fly so the model can attend to a larger canvas.
• SEGA – Spectral‑Energy Guided Attention, a frequency‑based sharpening alternative.
• SPA / HAP – HRDiT‑based patches that fix spatial disorder or prune attention heads for speed.
• PixelRush, FreeScale, HiFlow, I‑Max – cascade‑style up‑scalers that start from a base‑resolution image/latent and refine it progressively. All nodes are drop‑in: place the node after the model‑loader in a ComfyUI graph, set the target width/height, and connect it to the usual KSampler (or to the cascade‑specific inputs).

Key features

• Multi‑architecture support – works with Flux (including the quantised Nunchaku variant), Qwen‑Image 1.0 & 2.1, Krea‑2, Z‑Image, Anima/Cosmos and even classic SDXL/SD1.5 via the cascade nodes.
• Zero‑retraining – no extra checkpoints or fine‑tuning; the patch is applied at runtime.
• Single‑node integration – each method is a single ComfyUI node; you can mix‑and‑match (e.g., DyPE + SPA for quality, DyPE + HAP for speed).
• High‑resolution output – demonstrated up to 4096 × 4096 pixels and higher.
• Low overhead – DyPE/SEGA add negligible cost; SPA adds ~1.3‑1.8× inference time at 4 K; HAP can give 2‑5× speed‑ups when a scope‑plan is used.

When to use which node

Goal	Recommended node(s)
Generate directly at 4 K from noise	DyPE (or SEGA) – choose a method (vision_yarn, yarn, ntk, …) based on model and visual preference.
Fix repeated structures that appear when scaling	SPA (HRDiT) – adds a cheap spatial‑alignment pass.
Speed up high‑res sampling	HAP (HRDiT) – use a pre‑calibrated scope plan (provided for Flux) or calibrate your own.
Upscale an existing image while keeping its composition	PixelRush – patch‑wise cascade refinement.
Fully automatic upscale without a source image	FreeScale – self‑cascade with scale‑fused attention.
Upscale flow‑based models (Flux, Qwen‑Image, …) preserving structure	HiFlow – re‑uses the low‑res diffusion trajectory as guidance.
Dual‑pass high‑res for Flux/Z‑Image	I‑Max – low‑res pass guides a separate high‑res pass.

Getting started

1. Install ComfyUI (if you haven’t already).
2. Clone the repository into ComfyUI/custom_nodes/:

   cd path/to/ComfyUI/custom_nodes
   git clone https://github.com/wildminder/ComfyUI-DyPE.git
   

3. Restart ComfyUI – the new nodes appear under WMNodes → image.
4. Build a simple graph: Load Model → DyPE → KSampler → VAE Decode.
5. Set width/height to the desired high resolution (e.g., 4096 × 4096) and pick a method (the README recommends vision_yarn as a safe default).
6. Run the workflow.
   Optional: add SPA after DyPE if you notice duplicated patterns, or add HAP for a speed boost.

Limitations & gotchas

• The patches are model‑specific; some nodes (e.g., HAP) are not supported on every architecture (Nunchaku, Qwen‑Image 2.1, etc.).
• Qwen‑Image 2.1 uses a 16× VAE with an alpha channel; cascade nodes drop the alpha channel and log a warning.
• DyPE/SEGA require the correct model_type selection; auto‑detect works in most cases but manual override (flux, qwen21, …) may be needed for edge cases.
• High‑resolution runs still need sufficient GPU VRAM; the cascade nodes keep memory flat by processing patches, but the base model still occupies its full latent size.
• SPA and DyPE/SEGA cannot be used together – they are mutually exclusive.

Where to learn more

• Original papers linked in the README (DyPE, HRDiT, FreeScale, HiFlow) explain the underlying algorithms.
• Each node’s tooltip in ComfyUI shows the full list of parameters; the README provides recommended defaults and tuning tips.
• For custom speed‑up plans, see the calibration/calibrate_hap.py script.

---

ComfyUI‑DyPE is a practical toolbox for anyone who wants to generate 4 K‑plus images with the latest diffusion transformers while staying inside the visual‑workflow paradigm of ComfyUI.