atonamy/wan-alpha

Wan-Alpha text-to-video generation with alpha channel, optimized for Replicate deployment. WebM/WebP output with transparency support.

Public
0 runs
GitHub
Playground Examples README Versions

Wan-Alpha - Text-to-Video with Alpha Channel

Generate high-quality videos with transparent backgrounds using AI. Perfect for creating animations, visual effects, and compositing.

What is this?

Wan-Alpha generates videos from text descriptions with full alpha channel support - meaning transparent backgrounds you can overlay on anything.

Example: - Input: “A colorful parrot flying. Transparent background. Close-up shot. Realistic style.” - Output: WebM/WebP video with the parrot on a transparent background

Features

  • ✨ Transparent backgrounds - True alpha channel support
  • 🎬 WebM/WebP output - Industry-standard formats with transparency
  • 📐 Multiple resolutions - Including custom 512x512 square format
  • ⚡ Adjustable FPS - 1-60 frames per second
  • 💎 Quality control - Fine-tune output quality (1-100)

How to Use

Basic Usage

  import replicate

  output = replicate.run(
      "atonamy/wan-alpha",
      input={
          "prompt": "A butterfly flying through flowers. Transparent background. Medium shot.
  Realistic style.",
          "output_format": "webm",
          "resolution": "832*480",
          "num_frames": 81,
          "fps": 16,
          "output_quality": 85
      }
  )

print(output)  # URL to your generated video

Parameters

Required: - prompt (string) - Describe what you want to see. Important: Include “transparent background” and specify shot type (close-up, medium shot, wide shot)

Optional: - output_format (string) - webm or webp (default: webm) - resolution (string) - Options: - 480832 - Portrait - 832480 - Landscape (default) - 7201280 - Portrait HD - 1280720 - Landscape HD - 512512 (fit vertical) - Square, portrait-oriented - 512512 (fit horizontal) - Square, landscape-oriented - num_frames (int) - Number of frames, must be 4n+1 (e.g., 33, 49, 81). Default: 81 (~5 seconds at 16fps) - fps (int) - Frame rate 1-60. Default: 16 - output_quality (int) - Quality 1-100. Default: 85 - sample_steps (int) - Diffusion steps. Default: 4 - guidance_scale (float) - CFG scale. Default: 1.0 - seed (int) - Random seed. -1 for random. Default: -1 - negative_prompt (string) - What to avoid in the generation

Prompt Writing Tips

Good prompts include: 1. “Transparent background” - Essential for transparency 2. Shot type - “Close-up shot”, “Medium shot”, “Wide shot” 3. Subject description - What you want to see 4. Style - “Realistic style”, “Animated style”, etc.

Example prompts:

“A golden dragon flying. Transparent background. Wide shot. Realistic style.”

“Glowing particles floating upward. Transparent background. Close-up. Magical style.”

“A person waving hello. Transparent background. Medium shot. Realistic style.”

Advanced Example

  import replicate

  # Generate high-quality square video
  output = replicate.run(
      "atonamy/wan-alpha",
      input={
          "prompt": "Colorful smoke swirling. Transparent background. Close-up shot. Realistic 
  style.",
          "output_format": "webm",
          "resolution": "512*512 (fit vertical)",
          "num_frames": 49,  # ~3 seconds
          "fps": 24,  # Smoother motion
          "output_quality": 95,  # Higher quality
          "seed": 42  # Reproducible results
      }
  )

Output

  • Format: WebM or WebP file
  • Alpha channel: Full transparency support
  • Duration: Depends on num_frames and fps
    • 81 frames @ 16fps = ~5 seconds
    • 49 frames @ 24fps = ~2 seconds

Use Cases

  • 🎥 Video compositing and VFX
  • 🎮 Game assets and animations
  • 📱 AR filters and effects
  • 🎨 Motion graphics and overlays
  • 📺 Broadcast graphics
  • 🌐 Web animations

Technical Details

  • Model: Wan2.1-T2V-14B with alpha channel adaptation
  • Base resolution: Renders at native supported sizes, then optionally resizes
  • 512x512 formats: Rendered at higher resolution (480x832 or 832x480) then scaled down with transparent padding for best quality

Limitations

  • Requires “transparent background” in prompt for best results
  • Best with simple subjects on transparent backgrounds
  • Complex scenes may have edge artifacts
  • Frame count must follow 4n+1 pattern (33, 49, 81, etc.)

Credits

Based on https://github.com/WeChatCV/Wan-Alpha by the WeChatCV team.