Generate videos from text prompts with Wan2.2 on Deadline Cloud

The wan22_video_generation job bundle on the GitHub website generates short video clips from a text prompt using Wan2.2, an open video generation model. Each task renders one independent clip with its own seed, so a single job fans out across workers and returns a set of variations on the same prompt.

The bundle runs the Wan2.2 TI2V-5B checkpoint through the diffusers WanPipeline. Workers pull the weights from Hugging Face at runtime and cache them, preferring the fleet's persistent volume so that later workers reuse the download. No Hugging Face token is required because the Wan2.2 repositories are ungated. For the model itself, see the Wan2.2 repository on the GitHub website.

The bundle requires a farm with a GPU-enabled Linux queue. Workers must have an NVIDIA GPU (24 GB VRAM minimum), 64 GiB of system memory, and at least 60 GiB of free disk for the model cache. On 24 GB cards the script turns on sequential CPU offload and VAE tiling so the model fits; larger cards run considerably faster.

Submit a job that generates four variations of a prompt:

deadline bundle submit ./wan22_video_generation \
  --queue-id gpu-queue-id \
  -p Prompt="A hot air balloon drifting over terraced rice fields at dawn" \
  -p NumClips=4 \
  -p OutputDir=~/wan22_output

The README describes the full parameter set, including resolution and frame-count constraints, faster smoke-test settings, and measured generation times. For related generative AI examples, see Generate images in batch with a diffusion model on Deadline Cloud and Train and use a FLUX.2 Klein LoRA on Deadline Cloud. For persistent volume setup, see the fleet storage documentation linked from the README.