MAGREF/README.md

metadata

license: openrail++
tags:
  - image-to-video
  - wan2.1
  - finetune
  - magref
model_type: image-to-video
base_model: Wan2_1-I2V

MAGREF‑Video 14B I2V (WanGP-Ready)

📁 This repo bundles the FP16 MAGREF-14B Image‑to‑Video model with an INT8 quantized version optimized for WanGP.

---

📥 Files

• Wan2_1-Wan-I2V-MAGREF-14B_fp16_pure.safetensors – original Wan2.1 FP16 MAGREF model, cleaned for WanGP loading.
• Wan2_1-Wan-I2V-MAGREF-14B_quanto_fp16_int8.safetensors – INT8 quantized version for faster inference, same architecture.

These weights are based on the fp8 MAGREF checkpoint published by Kijai at WanVideo_comfy.

---

⚙️ Features

• Image‑to‑Video pipeline (i2v backbone)
• Optimized for prompt-based subject preservation
• INT8 quantized version included for memory efficiency

---

🛠 Installation & Usage (WanGP)

1. Clone/download this repo.

2. Ensure both .safetensors files are in the same directory.

3. Use this JSON as your finetune definition:

   {
     "model": {
       "name": "MAGREF‑14B I2V",
       "architecture": "i2v",
       "modules": [],
       "URLs": [
         "Wan2_1-Wan-I2V-MAGREF-14B_fp16_pure.safetensors",
         "Wan2_1-Wan-I2V-MAGREF-14B_quanto_fp16_int8.safetensors"
       ],
       "auto_quantize": false
     },
     "multi_images_gen_type": 1,
     "prompt_enhancer": ""
   }
   

4. Start WanGP with UI setup:

   python wgp.py --multiple-images
   

5. Choose “MAGREF‑14B I2V”, upload reference image(s), then generate!

---

📊 Model Info (Wan 2.1, MAGREF)

• MAGREF extends the Wan2.1 14B I2V with identity-preserving video generation—originally published by Kijai.
• Built on the Wan 2.1 framework, a state-of-the-art open-source video foundation model with a 14B parameter backbone.
• MAGREF specializes in maintaining consistent subject appearance across varied motions, automatically managing pose and depth.

---

🧠 Tips & Tricks

• 📸 Provide clear reference images—MAGREF handles blending but quality matters.
• 🎯 Use prompt slots like: “**[Left: ref‑image 1 with X], [Right: ref‑image 2 with Y]**” for layered scene control.
• 🎥 Fine-tune motion through clip speed (slower = softer movement).
• ⚡ FP16 is slow; use INT8 quant for real-time generation.

---

🧩 Credits

• MAGREF 14B I2V checkpoint by Kijai / WanVideo_comfy
• Wan 2.1 open-source video model framework
• WanGP UI integration for finetune loading

---

Download, drop in WanGP, and go wild with your characters in motion!