Wan 2.2: State-of-the-Art Local Video

Why Wan 2.2 Is the Current Standard

Wan 2.2 from Alibaba is, as of mid-2026, the most widely used local video generation model in ComfyUI. It is natively supported without any custom node packages (ComfyUI version 0.3.5 and later), it comes with five built-in workflow templates, and the 14B parameter version at FP8 precision produces genuinely high-quality video that holds up to professional review.

What makes Wan 2.2 stand out is the combination of quality at the 14B scale and the accessibility of the 1.3B model for users with limited VRAM. Both variants use the same node structure — you simply load a different model file. This means you can prototype on the 1.3B model and switch to the 14B for final renders without rebuilding your workflow.

Model Variants and VRAM Requirements

• Wan 2.2 T2V 1.3B: text-to-video, 1.3 billion parameters. Runs comfortably on 8 GB VRAM. Quality is noticeably lower than 14B but useful for rapid iteration and lower-end hardware.
• Wan 2.2 T2V 14B (FP8): text-to-video, 14 billion parameters at 8-bit float precision. Requires 14–16 GB VRAM. Best quality for text-to-video among local models.
• Wan 2.2 I2V 14B: image-to-video variant. Same VRAM requirement as T2V 14B. Animates a still image with temporal coherence (covered in Lesson 4).
• Wan 2.2 FLF2V: first-last-frame-to-video. Takes a start frame and end frame and generates the transition between them.

Download model files from Hugging Face (search "Wan-AI/Wan2.2"). Place downloaded safetensors files in ComfyUI/models/diffusion_models/ (for the main model) and the VAE file in ComfyUI/models/vae/. The CLIP/text encoder files go in ComfyUI/models/clip/.

Loading the Built-in Workflow

ComfyUI includes Wan 2.2 workflow templates built-in. Access them two ways:

• Workflow Templates panel: click the workflow menu in the top-left of ComfyUI and select "Browse Templates." Find the Wan 2.2 section. Click the T2V template to load it directly onto your canvas.
• Official docs: docs.comfy.org/tutorials/video/wan/wan2_2 has downloadable workflow JSON files. Drag any JSON onto the canvas to load it.

The built-in T2V workflow contains the correct nodes pre-wired: WanVideo Loader → CLIP Text Encode (Wan) → WanVideo Sampler → WanVideo Decode → Video Combine.

The Wan 2.2 Node Chain

Understanding the five core nodes helps you customize the workflow:

• WanVideo Loader: loads the Wan model file. Select your .safetensors file and the matching VAE. Set dtype to fp8_e4m3fn for FP8 precision.
• CLIP Text Encode (Wan): Wan uses its own text encoder, separate from standard CLIP. Do not use a generic CLIP Text Encode node — Wan's encoder is trained differently and produces better results with the model's conditioning system.
• WanVideo Sampler: the core generation node. Parameters: steps (30 default), cfg (6.0 default), width/height, num_frames (number of video frames to generate), shift (5.0 default).
• WanVideo Decode: decodes the latent video to pixel frames.
• Video Combine: from Video Helper Suite; assembles frames into MP4. Set frame_rate to 16 or 24 FPS.

Prompting Wan 2.2

Wan 2.2 responds best to natural language descriptions of the scene, motion, and cinematic qualities — similar to FLUX prompting rather than tag-based SD 1.5 prompting. Be specific about what is moving, how it moves, and the camera behavior:

• Good: "A woman walks along a rainy Tokyo street at night, neon signs reflected in puddles, camera tracking her from the side at medium distance, slow motion, cinematic color grading."
• Weak: "woman, rain, Tokyo, neon, cinematic"

Wan 2.2 does not benefit from negative prompts in the same way SD 1.5 does. A short negative prompt ("blurry, low quality, watermark") is optional but rarely changes output significantly at the 14B scale.

For clip length: num_frames of 81 at 16 FPS produces approximately 5 seconds of video. 161 frames gives about 10 seconds. Longer clips require substantially more VRAM and generation time.