hdtrnk/VACEPhantom

📖 Wan2.1 VACE + Phantom (Finetune)

Author / Creator: Inner_Reflections_AI
Original Guide: Wan VACE + Phantom Merge – An Inner Reflections Guide

---

🔹 About This Finetune

A regular VACE + Phantom merge (non‑Causvid) prepared for WanGP.
Converted to pure FP16 for reliable loading and optional INT8 quantization.

• Architecture: vace_14B
• Mode: Image/Video conditioning with multi‑image reference support (2–4 refs in custom WanGP builds)
• Variants: FP16 (pure) and quanto INT8

---

🔹 Files

• Wan2.1VACE_Phantom_fp16_pure.safetensors
• Wan2.1VACE_Phantom_quanto_fp16_int8.safetensors (or _quanto_bf16_int8 depending on your dtype selection in WanGP)

Replace these with your final filenames/links if different.

---

🔹 Usage in WanGP

Place the finetune JSON in:

app/finetunes/vace_phantom.json

Example JSON (matching the regular VACE + Phantom merge):

{
  "model": {
    "name": "VACE Phantom 14B",
    "architecture": "vace_14B",
    "description": "Regular VACE + Phantom merge by Inner_Reflections_AI, purified for WanGP. Multi-image references supported.",
    "URLs": [
      "ckpts/Wan2.1VACE_Phantom_fp16_pure.safetensors",
      "ckpts/Wan2.1VACE_Phantom_quanto_fp16_int8.safetensors"
    ],
    "modules": [],
    "auto_quantize": false
  }
}

---

🔹 Notes

• This is an experimental finetune. Tune steps, guidance scale, and reference image setup to taste.
• If you see a Gradio dropdown error (Value: on is not in the list...), refresh the UI and reselect the option.

---

🔹 Credits

• Merge & Guide: Inner_Reflections_AI
• WanGP Packaging: conversion to FP16 and finetune JSON layout compatible with WanGP.