README.md

metadata

title: Wan 2.2 Extend
emoji: 🎬
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 5.44.1
app_file: app.py
pinned: false
license: apache-2.0
short_description: Wan 2.2 I2V + Extend chaining for ~10–20s clips
hf_oauth: true
hf_oauth_scopes:
  - inference-api

Wan 2.2 Image → Video + Extend

Helper Space by Casey (Simzy): generate a short Wan 2.2 image-to-video clip on free ZeroGPU, then Extend by chaining until you reach roughly 10–20 seconds.

This Space does not load the 14B Wan weights itself. It calls the public ZeroGPU Space kulkas2pintu/wan222 via gradio_client, then stitches segments with ffmpeg.

How to use

0. Click Sign in with Hugging Face. Generation uses your own ZeroGPU quota; the Space owner's token is never used, and signed-out requests are refused.
1. Upload a still image and write a Generate prompt.
2. Click Generate — wait for ZeroGPU (often 1–3 minutes when busy).
3. Optionally set an Extend prompt (leave blank to reuse Generate prompt) and choose Extend start:
   • Last frame from video (default) — current behavior
   • Upload custom image — use a different image as the I2V start for the next segment
4. Click Extend to continue, or set a target length and use Auto-extend (capped at 6 segments).
5. Download the concatenated MP4.

Length guide

Segment length	Segments	Rough total
~3.5s	3	~10s
~3.5s	4–5	~14–18s
~4–5s	4	~16–20s

Honesty about ZeroGPU

• Each visitor signs in with HF OAuth; upstream calls to kulkas2pintu/wan222 carry that visitor's token, so ZeroGPU time and any overage bill to the visitor, not the Space owner. The Space has no HF_TOKEN secret.
• ZeroGPU is free but queued and quota-limited — not unlimited.
• Each segment is a separate upstream call; longer videos = more waits.
• If the queue is full or quota exhausted, wait and retry.

Architecture

[Your image] → Generate → clip1.mp4
                              ↓ last frame (or custom upload)
                         Extend → clip2.mp4 → ffmpeg concat → longer.mp4
                              ↓ …

Upstream API: generate_video on kulkas2pintu/wan222 (same endpoint used by Command Center).

Local files

Source also lives at /workspace/wan22-extend/ on the agent box for re-push if needed. Safe snapshot of v1: /workspace/wan22-extend-snapshots/v1-working-20260922-055108/ (+ .tgz). HF branch backup-v1-working is frozen and should not be overwritten.

License / ToS

Apache-2.0 for this app code. Hugging Face Terms of Service apply to hosted Spaces and ZeroGPU usage.