# Hosting Time-to-Move on Hugging Face

This guide explains how to mirror the Time-to-Move (TTM) codebase on the 🤗 Hub and how to expose an interactive demo through a Space. It assumes you already read the main `README.md`, understand how to run `run_wan.py`, and have access to Wan 2.2 weights through your Hugging Face account.

---

## 1. Prerequisites

- Hugging Face account with access to the Wan 2.2 Image-to-Video model (`Wan-AI/Wan2.2-I2V-A14B-Diffusers` at the time of writing).
- Local environment with Git, Git LFS, Python 3.10+, and the `huggingface_hub` CLI.
- GPU-backed hardware both locally (for testing) and on Spaces (A100 or A10 is strongly recommended; CPU-only tiers are too slow for Wan 2.2).
- Optional: organization namespace on the Hugging Face Hub (recommended if you want to publish under a team/org).

Authenticate once locally (this stores a token in `~/.huggingface`):

```bash
huggingface-cli login
git lfs install
```

---

## 2. Publish the code as a model repository

1. **Create an empty repo on the Hub.** Example:

   ```bash
   huggingface-cli repo create time-to-move/wan-ttm --type=model --yes
   git clone https://huggingface.co/time-to-move/wan-ttm
   cd wan-ttm
   ```

2. **Copy the TTM sources.** From the project root, copy the files that users need to reproduce inference:

   ```bash
   rsync -av \
     --exclude ".git/" \
     --exclude "outputs/" \
     /path/to/TTM/ \
     /path/to/wan-ttm/
   ```

   Make sure `pipelines/`, `run_wan.py`, `run_cog.py`, `run_svd.py`, `examples/`, and the new `huggingface_space/` folder are included. Track large binary assets:

   ```bash
   git lfs track "*.mp4" "*.png" "*.gif"
   git add .gitattributes
   ```

3. **Add a model card.** Reuse the main `README.md` or create a shorter version describing:
   - What Time-to-Move does.
   - How to run `run_wan.py` with the `motion_signal` + `mask`.
   - Which base model checkpoint the repo expects (Wan 2.2 I2V A14B).

4. **Push to the Hub.**

   ```bash
   git add .
   git commit -m "Initial commit of Time-to-Move Wan implementation"
   git push
   ```

Users can now do:

```python
from huggingface_hub import snapshot_download
snapshot_download("time-to-move/wan-ttm")
```

---

## 3. Prepare a Hugging Face Space (Gradio)

This repository now contains `huggingface_space/`, a ready-to-use Space template:

```
huggingface_space/
├── README.md                # Quickstart instructions
├── app.py                   # Gradio UI (loads Wan + Time-to-Move logic)
└── requirements.txt         # Runtime dependencies
```

### 3.1 Create the Space

```bash
huggingface-cli repo create time-to-move/wan-ttm-demo --type=space --sdk=gradio --yes
git clone https://huggingface.co/spaces/time-to-move/wan-ttm-demo
cd wan-ttm-demo
```

Copy everything from `huggingface_space/` into the Space repository (or keep the whole repo and set the Space’s working directory accordingly). Commit and push.

### 3.2 Configure hardware and secrets

- **Hardware:** Select an A100 (preferred) or A10 GPU runtime in the Space settings. Wan 2.2 is too heavy for CPUs.
- **WAN_MODEL_ID:** If you mirrored Wan 2.2 into your organization, set the environment variable to point to it. Otherwise leave the default (`Wan-AI/Wan2.2-I2V-A14B-Diffusers`).
- **HF_TOKEN / WAN_ACCESS_TOKEN:** Add a Space secret only if the Wan checkpoint is private. The Gradio app reads from `HF_TOKEN` automatically when calling `from_pretrained`.
- **PYTORCH_CUDA_ALLOC_CONF:** Recommended value `expandable_segments:True` to reduce CUDA fragmentation.

### 3.3 How the app works

`huggingface_space/app.py` exposes:

- A dropdown of the pre-packaged `examples/cutdrag_wan_*` prompts.
- Optional custom uploads (`first_frame`, `mask.mp4`, `motion_signal.mp4`) following the README workflow.
- Sliders for `tweak-index`, `tstrong-index`, guidance scale, seed, etc.
- Live status messages and a generated MP4 preview using `diffusers.utils.export_to_video`.

The UI lazily loads the `WanImageToVideoTTMPipeline` with tiling/slicing enabled to reduce VRAM usage. All preprocessing matches the logic in `run_wan.py` (the same `compute_hw_from_area` helper is reused).

If you need to customize the experience (e.g., restrict to certain prompts, enforce shorter sequences), edit `huggingface_space/app.py` before pushing.

---

## 4. Testing checklist

1. **Local dry-run.**
   ```bash
   pip install -r huggingface_space/requirements.txt
   WAN_MODEL_ID=Wan-AI/Wan2.2-I2V-A14B-Diffusers \
   python huggingface_space/app.py
   ```
   Ensure you can generate at least one of the bundled examples.

2. **Space smoke test.**
   - Open the deployed Space.
   - Run the default example (`cutdrag_wan_Monkey`) and confirm you receive a video in ~2–3 minutes on A100 hardware.
   - Optionally upload a small custom mask/video pair and verify that `tweak-index`/`tstrong-index` are honored.

3. **Monitor logs.** Use the Space “Logs” tab to confirm:
   - The pipeline downloads from the expected `WAN_MODEL_ID`.
   - VRAM usage stays within the selected hardware tier.

4. **Freeze dependencies.** When satisfied, tag the Space (`v1`, `demo`) so users know which TTM commit it matches.

You now have both a **model repository** (for anyone to clone/run) and a **public Space** for live demos. Feel free to adapt the instructions for the CogVideoX or Stable Video Diffusion pipelines if you plan to expose them as well; start by duplicating the provided Space template and swapping out `run_wan.py` for the relevant runner.