Video Generation - Wan2.2 TI2V 5B

🛠

Compact Model for Both Text-to-Video and Image-to-Video

Generate 720p video clips at 24 fps from a text prompt or from a text prompt anchored to a starting image using the same single model.

What it does

Wan2.2 TI2V 5B is a video generation model from the Wan-AI team. "TI2V" stands for Text-Image-to-Video, meaning the same model can do both text-to-video (give it just a text prompt and it imagines the scene from scratch) and image-to-video (give it a starting image plus a text prompt and it animates the image). Whichever mode you use depends only on whether you pass an image. The model outputs MP4 video at 720p (1280×704 or 704×1280) with 24 fps, and was designed specifically to be small and fast - only 5 billion parameters, runnable on a single consumer-grade GPU. Bilingual prompts (English and Chinese) are supported. The "secret sauce" behind its speed/size at this quality level is a high-compression video VAE (the Wan2.2-VAE) that shrinks video data much more aggressively than typical video diffusion VAEs while still reconstructing it cleanly.

Problem it solves

  • One model for two modes – Text-to-video and image-to-video in a single unified model, no need to swap checkpoints or pipelines
  • 720p / 24 fps output – Cinematic-class resolution and frame rate at small-model cost
  • Runs on consumer GPUs – Works on a single RTX 4090 (≥24 GB VRAM) — no datacenter GPU required
  • Fast – Among the fastest 720p@24fps open-source video generators; 5 seconds of 720p video in <9 minutes on a 4090
  • Animate a still image – Provide a reference image plus a text description; the model produces video starting from that image, following the prompt
  • Bilingual prompts – Natively supports both English and Chinese text prompts
  • Open and permissive – Apache 2.0 license; users own their generated content

Input/Output

  • Input:
    • Text to Video:
      • Plain-language description of the scene you want
      • English or Chinese
        • image
    • Image to Video:
      • One reference image to use as the first frame
      • A text description of what should happen starting from that image
  • Options (defaults from the HF model card):
    • Resolution (W × H, ratio) — default 1280 × 704 (16:9)
      • Pick the ratio that matches the input image (when using image-to-video)
    • Frame count (Wan VAE multiples) — default 121 frames (~5 s at 24 fps)
      • Must be a value the Wan2.2-VAE accepts — pick one of the offered multiples
    • Output FPS (24 is the native frame rate) — default 24
    • Diffusion steps (20 – 60, default 50)
      • More steps = higher quality, slower generation
    • Advanced Options:
      • CFG scale (default 5) — how strictly the output follows the prompt
      • Additional negative prompt (default empty) — extra description of what to avoid in the output. The model already ships with a built-in default negative prompt (Chinese-language list suppressing bad anatomy, extra fingers, JPEG artifacts, etc.); this field is for anything you want to add on top
      • Random seed (-1 = random, default 1) — set a fixed integer to reproduce the exact same video
  • Output: Generated video
    • Format: MP4
    • Resolution: 720p (1280×704 or 704×1280)
    • Frame rate: 24 fps
    • Length: Configurable via the frame count (default 121 frames ≈ 5 s at 24 fps)

Accuracy & Speed

Model size
5 billion parameters — a dense model (no Mixture-of-Experts)
Speed (single GPU)
A 5-second 720p video in < 9 minutes on a single consumer-grade GPU (e.g., RTX 4090); among the fastest 720p@24fps open-source video generators
VRAM requirement
≥ 24 GB VRAM for single-GPU 720p inference (e.g., RTX 4090)
Multi-GPU scaling
FSDP + DeepSpeed Ulysses for distributed inference; --ulysses_size 8 etc.
Bigger siblings
Wan2.2 T2V-A14B (text-to-video, MoE, ~27 B total / 14 B active) and Wan2.2 I2V-A14B (image-to-video, MoE) — higher quality, much heavier to run
Languages
English and Chinese prompts
Default video output
1280 × 704 at 24 fps, 121 frames (≈ 5 seconds)
Comparisons
The Wan team reports favorable results vs. leading closed-source commercial video generators on their Wan-Bench 2.0 evaluation

Technical Details

Architecture
Dense video diffusion transformer with the Wan2.2-VAE (a high-compression video VAE)
VAE Compression
4 × 16 × 16 (Time × Height × Width) — overall 64× compression vs. raw pixels
Effective Compression with Patchification
4 × 32 × 32 ≈ 4096× — what enables a 5B dense model to handle 720p video efficiently
Why it's small and fast
The high-compression VAE shrinks the latent space dramatically, so the diffusion transformer can work in a much smaller space — letting a 5B dense model match what would otherwise require a much bigger model
Model family note
The Wan2.2 family's larger 14B variants (T2V-A14B, I2V-A14B) use a Mixture-of-Experts architecture with a high-noise expert (early denoising) and a low-noise expert (later denoising). The 5B TI2V variant is dense — different design choice, optimized for compactness and speed
Parameters
5B (dense)
Precision
bfloat16 (recommended)
Default Inference Hyperparameters
num_inference_steps=50, guidance_scale=5.0, num_frames=121, height=704, width=1280, fps=24
Training Data (vs. Wan2.1)
+65.6% more images, +83.2% more videos; meticulously curated aesthetic data with detailed labels for lighting, composition, contrast, color tone
Supported Frameworks
Hugging Face Diffusers (WanPipeline, AutoencoderKLWan, WanTransformer3DModel, UniPCMultistepScheduler), Safetensors, ComfyUI (official integration)
Diffusers Note
Currently requires the main branch of diffusers (not the latest PyPI release): pip install git+https://github.com/huggingface/diffusers
Release date
July 28, 2025
Paper
"Wan: Open and Advanced Large-Scale Video Generative Models" — arXiv:2503.20314

Compliance & Provenance

Provider
Alibaba (open weights)
Provider type
Specialized
License
Apache 2.0
EU AI Act risk class
Limited Risk
Art. 50 transparency
Required — outputs are marked. See AI Policy §2.
Region availability
Available globally
Training data summary
Provider model card / training data

For more on how we classify models and mark outputs, see our AI Policy.

Model Source