Vantage DyPE Node for ComfyUI

Dynamic Position Extrapolation for Ultra-High-Resolution Flux Models

---

🧠 Overview

Vantage DyPE brings DyPE (Dynamic Position Extrapolation) directly into ComfyUI, allowing Flux-based diffusion transformer models to generate native ultra-high-resolution images — up to 8K — without retraining or using any external upscalers.

DyPE dynamically modulates the model’s positional embeddings during denoising, preserving geometry, proportion, and texture fidelity even when rendering far beyond the model’s original training resolution.

⚡ Generate true 4K+ images natively in ComfyUI using your existing Flux models — fully training-free, stable, and VRAM-optimized.

---

✨ Features

• 🧩 Seamless Flux Model Patching (Krea, Kontext, Dev, etc.)
• ⚙️ Native High-Resolution Sampling — supports 2K, 4K, 8K outputs
• 🌀 Integrated DyPE Positional Modulation (Yarn / NTK / Base)
• 🧮 Smart Sigma (σ) Remapping for large token grids
• 🧠 Adaptive Precision switching (fp16 / bf16)
• 💾 GGUF / Quantized Model Support for low-VRAM setups
• ⚡ Performance Mode — optimized memory & compute speed
• 🔲 Empty Latent Generator — create quick zero-latent tensors for testing

---

🧰Installation

1️⃣ Locate Your Custom Nodes Folder

Navigate to your ComfyUI installation directory and open:

ComfyUI/custom_nodes/

2️⃣ Clone or Copy the Node

Clone this repo directly inside custom_nodes:

git clone https://github.com/vantagewithai/Vantage-DyPE.git

Or manually copy the Vantage-DyPE folder into:

ComfyUI/custom_nodes/Vantage-DyPE/

3️⃣ Restart ComfyUI

After installation, restart ComfyUI.
The node will appear under Vantage → Model → / Patches.

---

⚙️ Usage Guide

1. Load a Flux-based model (Flux Krea, Flux Kontext, Flux Dev, etc.) using a:

   • Diffusion Model Loader (for safetensors)
   • UNet Loader (for GGUF)

2. Model Format Support:

   • ✅ .safetensors (BF16 or FP8 scaled)
   • ✅ .gguf (quantized)

3. If Using GGUF Models:

   • Enable the UNet Loader Node.
   • Connect UNet Loader → LoRA Loader → Sampler.
   • Disable the Diffusion Loader Node.

4. Load Additional Models:

   • 🔥 Flux Turbo Alpha LoRA — enables 8-step fast sampling.
   • 🎨 VAE Model: sd-vae-ft-mse (recommended).
   • 🧠 Encoders: CLIP-L and T5-XXL (text conditioning).

5. Connect Model Output → Vantage DyPE Node.
   Configure:

   • method: Extrapolation method (yarn recommended)
   • enable_dype: Enable/disable DyPE
   • dype_exponent: Modulation strength (2.0 default)
   • base_shift / max_shift: Sigma remapping curve
   • adaptive_precision / performance_mode: Precision & VRAM options

6. Connect DyPE Output → KSampler → VAE Decode.

💡 Tip: Keep both width and height divisible by 64 for stable geometry and best Flux attention alignment.

---

📏 Recommended Resolutions

| Type | Resolution | Notes | |------|-------------|-------| | Square | 1024×1024 → 4096×4096 | Best structural accuracy | | Landscape | 1920×1080 / 3840×2160 | Cinematic / 4K UHD | | Portrait | 2160×3840 | Stable up to 3K vertical | | Ultra | 4096×4096+ | Requires ≥24GB VRAM |

---

⚙️ Recommended Settings

| Parameter | Recommended | Description | |------------|-------------|-------------| | method | yarn | Stable & accurate extrapolation | | dype_exponent | 2.0 | Balanced between sharpness & stability | | base_shift | 0.5 | Smooth low-resolution adaptation | | max_shift | 1.15 | Ideal for 4K generation | | performance_mode | off | Keep off for GGUF or quantized models | | adaptive_precision | on | Enables dynamic fp16/bf16 switching |

---

🧩 Example Workflow

1. Load model (Flux Krea / Kontext / Dev).
2. Add Flux Turbo Alpha LoRA for fast 8-step sampling.
3. Insert Vantage DyPE Node between model loader and sampler.
4. Configure resolution and settings.
5. Connect to sampler and VAE decode node.
6. Run generation and enjoy native 4K output — no upscaler needed!

---

📦 Recommended Dependencies

• ComfyUI (latest build)
• PyTorch ≥ 2.2.0 with CUDA support
• Flux model (Krea / Kontext / Dev)
• Flux Turbo Alpha LoRA (optional)
• VAE: sd-vae-ft-mse
• Encoders: CLIP-L + T5-XXL

---

🔗 Attribution & Credits

📚 Research & Core Algorithm

DyPE: Dynamic Position Extrapolation for Ultra High Resolution Diffusion
By Hebrew University of Jerusalem
👉 Official DyPE GitHub Repository

🧩 Code Reference

Some parts of this node’s implementation and integration logic are adapted from:
👉 ComfyUI-DyPE by wildminder

🎨 Flux Models

By Black Forest Labs:

• Flux Krea
• Flux Kontext

---

🧾 License

This project is released under the MIT License.
Attribution to the original DyPE research authors and ComfyUI-DyPE contributors is required in any forks, distributions, or derivative works.

---

📸 Example Results

| Resolution | Model | Steps | Description | |-------------|--------|--------|-------------| | 3072×3072 | Flux Krea + DyPE + Turbo LoRA | 8 | Realistic portrait with perfect tone balance | | 3840×2160 | Flux Krea + DyPE | 8 | Native UHD fidelity without upscaling | | 4096×4096 | Flux Kontext + DyPE | 12 | Stable ultra-detailed render |

---

🧩 Troubleshooting

| Issue | Possible Cause | Fix | |--------|----------------|-----| | Elongated / squashed images | Non-multiple of 64 resolution | Use width & height divisible by 64 | | GPU memory overflow | Too high resolution or FP32 mode | Enable performance mode or lower res | | Weak detail at edges | Low dype_exponent | Increase to 2.0–2.5 | | Over-sharpening | High dype_exponent or aggressive sigma shift | Reduce to 1.5–2.0 | | No output with GGUF | Incorrect loader wiring | Use UNet Loader → LoRA Loader path |

---

💬 Author

Created by Vantage with AI
🎥 YouTube Channel — tutorials, model guides, and diffusion workflows.

“Bringing true native 4K+ generation to ComfyUI.”
— Vantage DyPE Node, 2025