Wan2.1-T2V-1.3B-Diffusers

This is a Text-to-Video model with 1.3 billion parameters, developed for generating video from text prompts. The model is optimized for consumer-grade GPUs: it requires 8.19 GB of VRAM, and generating a 5-second video in 480p resolution takes ~4 minutes on an RTX 4090 GPU without optimization.

Key Features:

• High performance outperforms open-source and commercial solutions across metrics.
• Support for multiple tasks:
  • Video generation from text (Text-to-Video),
  • Video generation from images (Image-to-Video),
  • Video editing, audio synthesis for video, and text-to-image generation,
  • Text generation — the first video model supporting Chinese and English text synthesis in video.
• Efficient VAE — Wan-VAE enables encoding/decoding of Full HD video of any length while preserving temporal sequence, minimizing memory consumption.

Technical Details:
Generation:

• Supports 480p and 720p resolutions (stability is lower for 720p due to limited training data).
• Generation time: ~4 minutes per 5-second video (480p, without optimization).
• Supports the Diffusers library (Hugging Face) for inference.
• Single-GPU mode: works smoothly on RTX 4090 with the offload_model option.
• Multi-GPU mode: scalable via FSDP + xDiT USP.

Prohibited Use: Generating content that violates laws, infringes on rights, or spreads misinformation. The model is intended for research and creative projects, balancing performance and accessibility.

---

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

• UMT5 text encoder: ~6B parameters,
• Transformer: ~1.3B parameters,
• VAE: ~127M parameters.

Total: ~7B parameters

---

The 'VRAM requirements' line is calculated based on the size of the weights in the specified precision, with an additional +10% coefficient for inference:
  model_parameters_count * bytes_per_parameter * 1.1 / 1024^3 
where model_parameters_count is number of model parameters, bytes_per_parameter is number of bytes used to store one parameter. For 16 bit = 2, 8 bit = 1, 4 bit = 0.5.

Announce Date: 01.03.2025
Parameters: 1.3B
Context: 512
VRAM requirements: 3.7 GB using 4 bits quantization, 7.4 GB using 8 bits quantization, 14.8 GB using 16 bits quantization
Developer: Alibaba Wan Team
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore Wan2.1-T2V-1.3B-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.