Eric-Alice-T2V-ComfyUI-Wrapper

Eric_Alice_t2v

ComfyUI custom nodes for Alice T2V 14B MoE - an open-source text-to-video model by Mirage that achieves state-of-the-art quality in 4 inference steps via score-regularized consistency distillation (rCM).

---

Example Output

Click the image to watch the sample video on YouTube.

---

Example Workflow

📥 Download the example workflow JSON - Drag and drop into ComfyUI to get started quickly!

---

Nodes

| Node | Description | |------|-------------| | Alice T2V Loader (Eric) | Loads and caches the Alice pipeline (T5 + 2× DiT + VAE) | | Alice T2V Generator (Eric) | Generates video frames from a text prompt |

---

Quick Start Setup

Follow these steps in order to get Alice T2V running in ComfyUI.

Step 1: Download the Model Weights from HuggingFace

The model weights (~27GB) are hosted on HuggingFace. Download them to a folder you'll remember:

Option A - Using huggingface-cli (recommended):

# Install huggingface-cli if you don't have it
pip install huggingface_hub

# Download to your preferred location (change the path as needed)
huggingface-cli download gomirageai/Alice-T2V-14B-MoE --local-dir "D:/models/Alice-T2V-14B-MoE"

Option B - Manual download:

1. Go to: https://huggingface.co/gomirageai/Alice-T2V-14B-MoE
2. Click "Files and versions"
3. Download all files into a folder, e.g. D:/models/Alice-T2V-14B-MoE

Important: Remember this folder path! You'll need to enter it in the Alice T2V Loader node's ckpt_dir field.

---

Step 2: Download the Alice Source Code from GitHub

The Alice Python code needs to be "vendored" (copied) into this node package. This is a one-time setup.

1. Download or clone the Alice repo:

   • Download ZIP: Go to https://github.com/mirage-video/Alice → Click green "Code" button → "Download ZIP" → Extract somewhere (e.g. C:/temp/Alice-main)
   • Or clone with git:

     git clone https://github.com/mirage-video/Alice.git C:/temp/Alice
     

2. Note the path where you extracted/cloned it (e.g. C:/temp/Alice or C:/temp/Alice-main)

---

Step 3: Run the Vendor Setup Script

This copies the necessary Python files into the node's vendor/ folder.

1. Open a terminal/command prompt

2. Navigate to this node's folder:

   cd ComfyUI/custom_nodes/Eric_Alice_t2v
   

3. Run the setup script, pointing to where you downloaded Alice:

   python setup_vendor.py --alice-src "C:/temp/Alice"
   

   (Adjust the path to match where you extracted/cloned the Alice repo)

4. You should see:

   Copying ...alice/ -> ...vendor/alice
   ✓ Vendor install verified - alice imports OK.
   

Troubleshooting: If you see an error about missing alice subfolder, make sure you're pointing to the repo root (the folder containing alice/, scripts/, configs/, etc.), not a subfolder.

---

Step 4: (Optional) Delete the Downloaded Alice Repo

The setup script copied everything needed into vendor/alice/. You can now safely delete the GitHub repo you downloaded in Step 2 to free up disk space:

• Delete C:/temp/Alice (or wherever you extracted it)

The node will work without it.

---

Step 5: Install the Extra Dependency

Only one extra Python package is needed:

pip install easydict

All other dependencies (torch, transformers, diffusers, etc.) are already present in a standard ComfyUI installation.

---

Usage in ComfyUI

1. Add the Alice T2V Loader node

   • Set ckpt_dir to your HuggingFace model folder from Step 1
     (e.g. D:/models/Alice-T2V-14B-MoE)

2. Add the Alice T2V Generator node

   • Connect the pipeline output from the Loader to the Generator
   • Enter your text prompt

3. Save the output

   • Connect frames output to VHS_VideoCombine (VideoHelperSuite) or SaveAnimatedWEBP

Recommended settings

| Parameter | Value | Notes | |-----------|-------|-------| | steps | 4 | Distilled for 4 steps - fast and high quality | | cfg_high_noise | 4.0 | Structure phase guidance | | cfg_low_noise | 3.0 | Detail phase guidance | | shift | 8.0 | Timestep shift | | solver | unipc | Best for low step counts | | frame_num | 81 | 5 seconds at 16 fps | | offload_model | True | Safe default; disable on 96GB+ VRAM for speed | | t5_cpu | True | Keeps VRAM free for DiT experts |

---

VRAM requirements

| Config | Approx. VRAM | |--------|-------------| | offload_model=True, t5_cpu=True, convert_model_dtype=False | ~28 GB | | offload_model=False, t5_cpu=True | ~56 GB (both experts loaded) | | offload_model=False, t5_cpu=False | ~70+ GB |

---

Output format

frames is a standard ComfyUI IMAGE tensor: [T, H, W, C] float32 in [0, 1]. Plug directly into any node that accepts an IMAGE batch.

---

Links

| Resource | URL | |----------|-----| | Model Weights (HuggingFace) | https://huggingface.co/gomirageai/Alice-T2V-14B-MoE | | Alice Source Code (GitHub) | https://github.com/mirage-video/Alice | | Mirage AI | https://gomirage.ai |

---

License

This ComfyUI wrapper is released under the MIT License - see LICENSE for details.

Alice Model License

The Alice T2V model and source code by Mirage AI are licensed under the Apache License 2.0.

• Model: https://huggingface.co/gomirageai/Alice-T2V-14B-MoE
• Source: https://github.com/mirage-video/Alice

---

Credits

• Alice T2V model by Mirage AI - Apache 2.0 License
• ComfyUI wrapper by Eric Hiss ([email protected])

This is an unofficial community wrapper and is not affiliated with or endorsed by Mirage AI.