wan-video/wan-2.7-t2v

Generate videos with audio from text prompts using Alibaba's Wan 2.7 model. 1080p, up to 15 seconds, with audio synchronization.

33 runs
GitHub
Paper
Playground API Examples README

Wan 2.7 text-to-video

Wan 2.7 is a text-to-video model from Alibaba’s Wan family. Describe a scene in natural language and it generates a video with coherent motion, lighting, and synchronized audio.

The model is built on a 27 billion parameter Mixture-of-Experts architecture and generates up to 1080p video at durations from 2 to 15 seconds. It auto-generates matching audio (sound effects, ambient noise) or you can provide your own audio file for voice or music synchronization.

Inputs

  • prompt — Text description of the video to generate (required)
  • negative_prompt — Describes content that should not appear in the video
  • audio — Optional audio file (wav/mp3, 3–30s, ≤15 MB) for voice or music synchronization. If omitted, the model generates matching audio automatically.
  • resolution — 720p or 1080p (default: 1080p)
  • aspect_ratio — 16:9, 9:16, 1:1, 4:3, or 3:4 (default: 16:9)
  • duration — Length in seconds, 2–15 (default: 5)
  • enable_prompt_expansion — Automatically expand short prompts for better results. Improves quality but adds latency (default: true)
  • seed — Random seed for reproducible results

Tips

  • Be descriptive. Include details about the scene, lighting, camera movement, and action. “A golden retriever running through autumn leaves in a park, camera tracking from the side, warm afternoon light” works much better than “a dog in a park.”
  • Keep durations short. 2–5 second clips tend to produce the most coherent motion and scene consistency.
  • Use negative prompts to reduce common artifacts — try “blurry, distorted, low quality, static.”
  • Enable prompt expansion for short prompts. It fills in visual details that improve generation quality.
  • Pick the right aspect ratio for your use case — 9:16 for vertical/mobile content, 16:9 for widescreen, 1:1 for social.

Limitations

  • Complex multi-character scenes with specific interactions can be inconsistent.
  • Text rendering within generated videos is unreliable.
  • Longer durations (10+ seconds) may show motion degradation or scene drift.
  • Precise spatial relationships (“object A is to the left of object B”) are not always followed exactly.
Model created
Model updated