workflow-templates-broken-zap-2026-06-08/site/knowledge/models/wan.md

Wan

Wan is a family of open-source video generation models from Alibaba's Tongyi Lab, spanning text-to-video, image-to-video, speech-to-video, motion control, and video editing. All models are released under the Apache 2.0 license.

Model Variants

Wan 2.1 T2V / I2V

• Text-to-video and image-to-video generation
• Available in 1.3B and 14B parameter sizes
• Supports 480p and 720p output, variable aspect ratios
• Chinese and English visual text generation

Wan 2.1 Fun (Control / InPaint / Camera)

• Camera control with predefined or custom camera movements
• Video inpainting for targeted frame-level editing
• Depth, pose, and canny edge control for guided generation

Wan 2.1 VACE (Video Any-Condition Editing)

• All-in-one model for video creation and editing (ICCV 2025)
• Reference-to-video (R2V), video-to-video (V2V), and masked editing (MV2V)
• Supports inpainting, outpainting, first-last-frame interpolation, and animate-anything
• Available in 1.3B and 14B sizes, built on Wan 2.1 base models

Wan 2.2 T2V / I2V / TI2V

• Mixture-of-Experts (MoE) architecture with high-noise and low-noise expert models
• T2V-A14B and I2V-A14B (14B MoE), TI2V-5B (hybrid text+image-to-video)
• Cinematic-level aesthetic control with lighting, composition, and color tone guidance
• TI2V-5B uses a high-compression 16×16×4 VAE, runs on consumer GPUs like 4090

Wan 2.2 S2V (Speech-to-Video)

• Audio-driven cinematic video generation from image + speech + text
• Supports lip-sync, facial expressions, and pose-driven generation
• Generates variable-length videos matching input audio duration

Wan 2.2 Animate

• Character animation and subject replacement from video + reference image
• Animate mode: transfers motion from reference video onto a still character
• Replace mode: swaps subjects while preserving background, lighting, and camera motion
• Includes relighting LoRA for scene-matched lighting adaptation

Wan Move

• Point-level motion control for image-to-video generation (NeurIPS 2025)
• Dense trajectory-based guidance for fine-grained object motion
• Latent trajectory propagation without extra motion modules
• 14B model generating 5-second 480p videos

Key Features

• High temporal consistency and natural physics simulation
• Multiple aspect ratios (16:9, 9:16, 1:1) at 24fps
• MoE architecture in 2.2 for higher quality at same compute cost
• Bilingual prompt support (Chinese and English)
• Hanzo Studio and Diffusers integration across all variants

Hardware Requirements

• 1.3B models: 8GB VRAM minimum
• 14B models: 24GB+ VRAM recommended (80GB for full precision)
• TI2V-5B: runs on consumer 4090 GPUs at 720p
• FP8 quantization available for lower VRAM configurations
• Multi-GPU inference supported via FSDP + DeepSpeed Ulysses

Common Use Cases

• Social media and short-form video content
• Character animation and motion transfer
• Video inpainting and scene editing
• Product animation and marketing videos
• Speech-driven talking head generation
• Storyboard-to-video conversion

Key Parameters

• frames: Number of output frames (typically 81 for ~3.4s at 24fps)
• steps: Inference steps (20-50 recommended)
• cfg_scale: Guidance scale for prompt adherence (3-7 typical)
• size: Output resolution (480p or 720p)
• model_name: Selects variant (e.g., vace-14B, ti2v-5B, s2v-14B)

Blog References

• Wan 2.1 Video Model Native Support — Initial release with 4 model variants, 8.19GB VRAM minimum
• Wan 2.1 VACE Native Support — Unified video editing: Move/Swap/Reference/Expand/Animate Anything
• Wan 2.2 Day-0 Support — MoE architecture, Apache 2.0 license, cinematic controls
• WAN 2.6 Reference-to-Video — Generate videos from reference clips at up to 1080p
• The Complete AI Upscaling Handbook — Wan 2.2 used for creative video upscaling