hdtrnk / FusioniX-All

A complete collection of WanGP FusioniX checkpoints
– both pure 16-bit dumps and pre-quantized 8-bit versions for image-to-video (I2V) and Phantom text-to-video (T2V) pipelines.

---

📂 Files in this repo

Pure FP16 checkpoints (~28–32 GB each)

• Wan14Bi2vFusioniX_pure_fp16.safetensors
  Wan2.1 Image→Video FusioniX 14B, pure 16-bit

• Wan14BT2VFusioniX_Phantom_pure_fp16.safetensors
  Wan2.1 Phantom Text→Video FusioniX 14B, pure 16-bit

Pre-quantized BF16→Int8 checkpoints (~14–16 GB each)

• Wan14Bi2vFusioniX_pure_quanto_bf16_int8.safetensors
  I2V FusioniX merged, quantized via Quanto (bf16→int8)

• Wan14BT2VFusioniX_Phantom_pure_quanto_bf16_int8.safetensors
  Phantom T2V FusioniX merged, quantized via Quanto (bf16→int8)

---

🚀 Quick Start (WanGP)

1. Copy your desired file into your WanGP ckpts/ folder.
2. Edit your finetune JSON (in app/finetunes/) so that under "model":

   "architecture": "i2v",                // or "phantom_14B"
   "URLs": ["ckpts/<filename>.safetensors"],
   "auto_quantize": false
   

3. **Restart** WanGP (no `--save-quantized` flag).
4. In the web UI, set **Transformer Model Quantization → Scaled Int8** (for the int8 files) or leave at BF16/FP16 for pure-16-bit.
5. **Generate**!

---

## ⚙️ Recommended Defaults

### I2V FusioniX

```yaml
resolution:     "832x480"
video_length:   81
num_steps:      8    # sweet-spot: 6–10
guidance_scale: 6.0
flow_shift:     7.0
embed_guide:    6
prompt_enhancer: on
slg_switch:     0
```

### Phantom T2V FusioniX

```yaml
resolution:     "832x480"
video_length:   81
num_steps:      8    # sweet-spot: 6–10
guidance_scale: 7.5
flow_shift:     5.0
embed_guide:    6
prompt_enhancer: on
slg_switch:     0
```

---

## 🔍 Architecture & Compatibility

* **WanGP architectures**:

  * `i2v` for image-to-video
  * `phantom_14B` for Phantom text-to-video
  * (See `model_signatures` in `wgp.py` or browse `app/settings/*.json` for all supported keys.)

* **WanGP version**: ≥ v6.0 (mmgp 3.4.9)

* **ComfyUI**: You can load these safetensors in ComfyUI’s Checkpoint Merger or as a Diffusers pipeline, but you may need a custom video pipeline config. They’re *not* plug-and-play ComfyUI “modelcards.”

---

## 📋 JSON Snippets

### I2V Finetune JSON

```json
{
  "model": {
    "name": "Wan image2video FusioniX 14B",
    "architecture": "i2v",
    "URLs": ["ckpts/Wan14Bi2vFusioniX_pure_quanto_bf16_int8.safetensors"],
    "auto_quantize": false
  },
  "... other settings ..."
}
```

### Phantom T2V Finetune JSON

```json
{
  "model": {
    "name": "Wan text2video Phantom FusioniX 14B",
    "architecture": "phantom_14B",
    "URLs": ["ckpts/Wan14BT2VFusioniX_Phantom_pure_quanto_bf16_int8.safetensors"],
    "auto_quantize": false
  },
  "... other settings ..."
}
```

---

## 🔗 Upstream & Licensing

* **Original WanGP repo** by DeepBeepMeep:
  [https://github.com/DeepBeepMeep/WanGP](https://github.com/DeepBeepMeep/WanGP)

* **Original HF models**:
  [https://huggingface.co/DeepBeepMeep/Wan2.1](https://huggingface.co/DeepBeepMeep/Wan2.1)

* **License**: Apache-2.0 (inherited from upstream)

---

*Happy generating!* 🎥✨\`\`\`