Wan2.2-T2V-A14B-Diffusers

Architecture:

• Uses a Mixture-of-Experts (MoE) with two specialized experts (high and low noise levels) to optimize the diffusion process. The model has a total capacity of 27B parameters (equivalent to 14B active parameters per step).
• Training: Trained on expanded datasets: +65.6% more images and +83.2% more video data compared to the previous version, improving generation of dynamic scenes and aesthetics.
• Compression: Includes Wan2.2-VAE with a compression ratio of 4×16×16 for efficient storage and generation without quality loss.

Technical Details:

• Supported Tasks: Unified architecture for Text-to-Video (T2V) and Image-to-Video (I2V) generation.
• Resolution Support: 720P@24fps, capable of running on NVIDIA RTX 4090-level GPUs.
• Inference:
  • Utilizes FSDP (Fully Sharded Data Parallel) and DeepSpeed Ulysses for multi-GPU inference.
  • Supports CPU offload to reduce GPU load (parameters: --offload_model, --t5_cpu).
  • Integration with Diffusers: Accessible via WanPipeline and AutoencoderKLWan from the Diffusers library.
  • Single-GPU Mode: Requires a minimum of 80GB VRAM, with options --offload_model and --convert_model_dtype.
  • Multi-GPU Mode: Uses torchrun with --ulysses_size configuration and FSDP enabled for T5 and DIT modules.

Use of generated content is permitted provided that legal requirements are followed and no harmful applications are made. Cite this work when used in research.

---

The model is a component of the video generation pipeline, consisting of:

• UMT5 text encoder: ~6B parameters,
• Transformer: ~27B parameters,
• VAE: ~127M parameters.

Total: ~34B parameters

Announce Date: 28.07.2025
Parameters: 27B
Experts: 2
Activated at inference: 14B
Developer: Alibaba Wan Team
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore Wan2.2-T2V-A14B-Diffusers capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.