HY-WU (Part I): An Extensible Functional Neural Memory Framework and An Instantiation in Text-Guided Image Editing

🔥 News

• March 10, 2026: We are hiring Interns and Full-time Employees! 🚀 (Focus: Parameter Generation. Drop your CV via victorkwang@global.tencent.com)
• March 6, 2026: 🎉 HY-WU open source - Inference code and model weights publicly available.

🗂️ Contents

• 🔥 News
• 📖 Introduction
• ✨ Key Features
• 🖼 Showcases
• 📑 Open-Source Plan
• 🚀 Usage
• 🧱 Memory Requirement
• 📊 Evaluation
• 📚 Citation

---

📖 Introduction

We propose HY-WU: a scalable framework for on-the-fly conditional generation of low-rank (LoRA) updates. HY-WU synthesizes instance-conditioned adapter weights from hybrid image–instruction representations and injects them into a frozen backbone during the forward pass, producing instance-specific operators without test-time optimization.

✨ Key Features

• 🧠 Functional Neural Memory: HY-WU introduces a lightweight “neural memory” for AI. It generates conditioned model adapter per request (without finetuning!), enabling instance-level personalization while preserving the base model’s general capability.

• 🏆 Scalable for Large Models: HY-WU remains practical for large foundation models (even at 80B parameters!). With structured parameter tokenization, the method naturally compatible with large-scale architectures.

• 🎨 Strong Human Preference: HY-WU achieves high human preference win-rates against open-source models, exceeds strong closed-source baselines, and remains close to the latest Nano-Banana series.

🖼 Showcases

Showcase 1: Cross-Domain Clothing Fusion

Showcase 2: Creative Cosplay and Character Outfit Migration

Showcase 3: High-Fidelity Face Identity Transfer

Showcase 4: Seamless Outfit Transfer and Virtual Try-on

Showcase 5: High-Quality Texture Synthesis

📑 Open-source Plan

• HY-WU
  • Inference
  • HY-Image-3.0-Instruct's checkpoint
  • Distilled checkpoint
  • Other models' checkpoint

🚀 Usage

🏠 Clone the repository

git clone https://github.com/Tencent-Hunyuan/HY-WU.git
cd HY-WU

📥 Install dependencies

pip install -r requirements.txt

🔥 Play with the code

Directly run infer.py

python infer.py

Or use the code below:

from wu import WUPipeline

base_model_path = "tencent/HunyuanImage-3.0-Instruct"
pg_model_path = "tencent/HY-WU"

pipeline = WUPipeline(
    base_model_path=base_model_path,
    pg_model_path=pg_model_path,
    device_map="auto",
    moe_impl="eager",
    moe_drop_tokens=False,
)

prompt = "以图1为底图，将图2公仔穿的衣物换到图1人物身上；保持图1人物、姿态和背景不变，自然贴合并融合。"
# prompt_en = Using Figure 1 as the base image, replace the clothing on the character in Figure 1 with the outfit worn by the figurine in Figure 2. Keep the character, pose, and background of Figure 1 unchanged, ensuring the new clothing fits naturally and blends seamlessly.
imgs_input = ["./assets/input_1_1.png", "./assets/input_1_2.png"]

sample = pipeline.generate(prompt=prompt, imgs_input=imgs_input, diff_infer_steps=50, seed=42, verbose=2)

sample.save("./output.png")

🎨 Interactive Gradio Demo

Launch an interactive web interface for easy image-to-image generation.

pip install gradio>=4.21.0

python gradio/app.py

🌐 Web Interface: Open your browser and navigate to http://localhost:7680 or shared link.

🧱 Memory Requirement

Base model param	HY-WU param	Recommended VRAM
80B (13B active)	8B	≥ 8 × 40 GB or 4 x 80GB

Notes:

• Multi‑GPU inference is required for the base model.

📊 Evaluation

👥 GSB (Human Evaluation)

HY-WU substantially outperforms leading open-source models, and remain competitive with top-tier closed-source commercial systems. While Nano Banana 2 and Nano Banana Pro achieve slightly higher overall scores (52.4% and 53.8%, respectively), the margin remains modest.

Given that these commercial systems are likely trained with substantially larger-scale backbones and proprietary data, the modest performance gap suggests that our operator-level conditional adaptation remains effective even under more constrained model scale.

📚 Citation

If you find HY-WU useful in your research, please cite our work:

@article{wu2026hy-wu,
  title={HY-WU (Part I): An Extensible Functional Neural Memory Framework and An Instantiation in Text-Guided Image Editing},
  author={Tencent HY Team, Mengxuan Wu, Xuanlei Zhao, Ziqiao Wang, Ruicheng Feng, Atlas Wang, Qinglin Lu, and Kai Wang},
  journal={arXiv preprint arXiv:2603.07236},
  year={2026}
}