Wan 2.2 5B Video Generation
Fast text-to-video and image-to-video generation with 5B parameters. Optimized for rapid prototyping and creative exploration.
Try an example
About this workflow
This ComfyUI workflow demonstrates fast, high-quality text-to-video and image-to-video generation using the Wan 2.2 5B model. It wires together the essential building blocks: UNETLoader loads the Wan2.2 diffusion backbone, CLIPLoader and CLIPTextEncode turn your positive and negative prompts into conditioning, and ModelSamplingSD3 configures the sampler so KSampler can produce temporally consistent latents.
For pure text-to-video, the diffusion model starts from noise; for image-to-video, Wan22ImageToVideoLatent injects structure and motion cues from a source image to initialize or steer the sequence.
Once frames are synthesized in latent space, VAELoader brings in the Wan 2.2 VAE and VAEDecode turns latents into RGB frames. CreateVideo assembles those frames at your chosen FPS and length, and SaveVideo writes the final clip to disk. The workflow is organized into helpful groups: Step1 - Load models to point at wan2.2_ti2v_5B_fp16.safetensors and wan2.2_vae.safetensors, Step3 - Prompt to edit text conditioning, a Video Size & length group to set resolution and frame count, and a For i2v, use Ctrl + B to enable branch that converts a still image into a moving shot.
It’s optimized for rapid prototyping—tweak prompts, steps, and sizes, then iterate quickly to explore ideas.
Questions & answers
Place wan2.2_ti2v_5B_fp16.safetensors in ComfyUI/models/diffusion_models and wan2.2_vae.safetensors in ComfyUI/models/vae. In Step1 - Load models, point UNETLoader to the diffusion model and VAELoader to the VAE. CLIPLoader will load the paired text encoder.
For text-to-video, keep the i2v branch bypassed. For image-to-video, load your image into LoadImage, then go to the group labeled For i2v, use Ctrl + B to enable and press Ctrl+B to un-bypass that branch. Use Wan22ImageToVideoLatent’s strength to set how much the source image influences the motion and look.
Increase frames or adjust prompts to include motion cues (e.g., slow pan, gentle breeze). For i2v, raise or lower Wan22ImageToVideoLatent strength to balance faithfulness vs. motion. Try more sampling steps or tweak guidance (CFG) for better temporal coherence, and refine your negative prompt to suppress unwanted artifacts.
Start small and scale up: use lower resolution (e.g., 640×360), fewer frames, and moderate steps for quick drafts. Keep batch size at 1. Use the fp16 model provided. If you hit memory limits, further reduce width/height or frame count. Higher resolutions and longer clips increase VRAM and render time.
SaveVideo controls the output path and filename; open that node to see or change where the MP4 is written. The workflow outputs silent video—audio is not generated.