Repolode
Enviar
Wan-Video

Wan-Video/Wan2.2

Open-source large-scale video generative models for text-to-video, image-to-video, speech-to-video and character animation.

Alternativa open source a:RunwaySora

Ver no GitHubSite
Estrelas
18,011
Forks
2,335
Issues abertas
320
Último commit
há 28 dias

O que é?

Wan2.2 is a family of open video generation models from the Wan team. It covers text-to-video (T2V-A14B), image-to-video (I2V-A14B), a 5B text-image-to-video model (TI2V-5B) that targets 720P at 24fps, speech-to-video (S2V-14B) driven by audio, and Wan-Animate-14B, which animates a character from a reference image using a motion video or replaces the character in a video.

The A14B models use a Mixture-of-Experts design with a high-noise expert for early denoising steps and a low-noise expert for later refinement, giving about 27B total parameters with roughly 14B active per step. The TI2V-5B model uses a high-compression Wan2.2-VAE and can run on a single consumer GPU such as an RTX 4090 with offloading options.

The repository provides command-line inference scripts with single-GPU and multi-GPU (FSDP and DeepSpeed Ulysses) modes, optional prompt extension via the DashScope API or local Qwen models, and weights hosted on Hugging Face and ModelScope. Diffusers and ComfyUI integrations are also available.

Recursos principais

  • 01Text-to-video and image-to-video generation at 480P and 720P
  • 02TI2V-5B model that generates 720P video at 24fps on consumer GPUs
  • 03Speech-to-video generation driven by audio input, with optional CosyVoice text-to-speech
  • 04Character animation and replacement mode for Wan-Animate
  • 05Single-GPU and multi-GPU inference with FSDP and DeepSpeed Ulysses
  • 06Optional prompt extension via DashScope API or local Qwen models

Casos de uso

  • Generating short video clips from a text prompt for creative or research work
  • Animating a still image of a character using a reference motion video
  • Producing talking or singing videos driven by an audio track
  • Running video generation experiments on a single high-memory or consumer GPU

Tópicos

#video-generation#text-to-video#image-to-video#diffusion-model#mixture-of-experts#pytorch#speech-to-video#generative-ai#aigc

Perguntas frequentes

O que é Wan2.2?
Open-source large-scale video generative models for text-to-video, image-to-video, speech-to-video and character animation. Wan2.2 is a family of open video generation models from the Wan team. It covers text-to-video (T2V-A14B), image-to-video (I2V-A14B), a 5B text-image-to-video model (TI2V-5B) that targets 720P at 24fps, speech-to-video (S2V-14B) driven by audio, and Wan-Animate-14B, which animates a character from a reference image using a motion video or replaces the character in a video.
Wan2.2 é gratuito? Qual é a licença?
Wan2.2 é open source sob a licença Apache-2.0. Confira os termos da licença antes de usá-lo em produtos comerciais.
Wan2.2 ainda é mantido?
Manutenção ativa. O último commit de Wan2.2 foi em 21 de set. de 2026.
Para que posso usar Wan2.2?
Generating short video clips from a text prompt for creative or research work Animating a still image of a character using a reference motion video Producing talking or singing videos driven by an audio track Running video generation experiments on a single high-memory or consumer GPU
Onde consigo Wan2.2?
O código-fonte e as instruções de instalação estão no GitHub: https://github.com/Wan-Video/Wan2.2

Resumo gerado por IA a partir do README do projeto. As estatísticas do GitHub são atualizadas diariamente.

Projetos relacionados