Wan
An open-source video generation foundation model from Alibaba's Tongyi Lab, built on a Diffusion Transformer (DiT) architecture. It processes video by dividing frames into 16×16 pixel spatiotemporal patches.
Timeline (last 90 days)
标准文生视频扩散管线（如原版 Stable Video Diffusion 或早期 Wan checkpoint）在超过 4 秒的帧序列上会出现时间注意力衰减。
ReelMint offers a separate Private Studio plan at $999 per month, whose generated content does not go through platform screening, supports Wan and Seedream models at up to 4K resolution, and can generate ten works at a time.
Seedance, Kling, Wan, Veo, Grok Imagine, MiniMax H3 and others are currently mainstream video generation foundation models, trained by different vendors and offered externally via APIs.
Mainstream T2V models include the Runway Gen series, Sora, Kling, and Wan.
Community comments suggest that the Qwen team and the Wan team have shortcomings in the quality of their real-world datasets.
Mainstream open-source video models such as CogVideoX and Wan generally suffer from imprecise responses in camera control.
Acceleration LoRA technology has been widely applied to accelerate the local deployment of mainstream open-source video models such as Wan and HunyuanVideo.
Compared to contemporaneous models such as Wan and CogVideoX, the MiniMax model has a larger parameter scale, placing higher hardware requirements on local deployment.
Mainstream locally deployable video generation models include ByteDance's CogVideoX and Kuaishou's Wan, while on the commercial API side there are Runway, Pika, Kling, and others.
The creator still retains a collection of LoRAs accumulated from previous video generation with Wan, believing that some of them are irreplaceable.
3 more timeline events
All Facts (13)
标准文生视频扩散管线（如原版 Stable Video Diffusion 或早期 Wan checkpoint）在超过 4 秒的帧序列上会出现时间注意力衰减。
50%UnverifiedReelMint offers a separate Private Studio plan at $999 per month, whose generated content does not go through platform screening, supports Wan and Seedream models at up to 4K resolution, and can generate ten works at a time.
50%UnverifiedSeedance, Kling, Wan, Veo, Grok Imagine, MiniMax H3 and others are currently mainstream video generation foundation models, trained by different vendors and offered externally via APIs.
50%UnverifiedMainstream T2V models include the Runway Gen series, Sora, Kling, and Wan.
50%UnverifiedCommunity comments suggest that the Qwen team and the Wan team have shortcomings in the quality of their real-world datasets.
50%UnverifiedMainstream open-source video models such as CogVideoX and Wan generally suffer from imprecise responses in camera control.
50%UnverifiedAcceleration LoRA technology has been widely applied to accelerate the local deployment of mainstream open-source video models such as Wan and HunyuanVideo.
50%UnverifiedCompared to contemporaneous models such as Wan and CogVideoX, the MiniMax model has a larger parameter scale, placing higher hardware requirements on local deployment.
50%UnverifiedMainstream locally deployable video generation models include ByteDance's CogVideoX and Kuaishou's Wan, while on the commercial API side there are Runway, Pika, Kling, and others.
50%UnverifiedThe creator still retains a collection of LoRAs accumulated from previous video generation with Wan, believing that some of them are irreplaceable.
50%UnverifiedUsing the ti2v-5b Q5 model at 960x544 resolution, 81 frames, and 20 steps, each step takes only 1.95 seconds, with approximately 3GB of VRAM remaining idle.
50%UnverifiedThe Wan series released by Alibaba's Tongyi team was once the mainstream choice for open-source video generation.
50%UnverifiedThe Wan team has open-sourced the character animation model Wan-Animate-2.
50%