.ai

the best open source ai models

Guides

Rankings

FAQ

Upload

WanImageToVideo and Wan 2.x video nodes node in ComfyUI

Q: Out of memory at the VAE decode

Decoding 81 frames at 720p in one go. Use VAE Decode (Tiled) with a temporal tile, lower the resolution, or decode at 480p and upscale frames afterwards.

Q: Load CLIP fails with the umt5 file

The type dropdown on Load CLIP is not set to wan. Set type to wan (and use the fp8 umt5_xxl file to save memory).

WanImageToVideo builds the conditioning and the empty video latent for a Wan 2.1/2.2 image-to-video render from a start image, a CLIP Vision embedding and the frame count; the surrounding graph uses Load Diffusion Model, Load CLIP, Load VAE and KSampler.

Node pack: ComfyUI core

Category: Video and animation

What WanImageToVideo and Wan 2.x video nodes does

Wan 2.x (Alibaba) is the most used open video model in ComfyUI. The core nodes WanImageToVideo (I2V), WanFunControlToVideo, WanFunInpaintToVideo, WanVaceToVideo (VACE control and reference) and WanFirstLastFrameToVideo take the positive and negative conditioning, the VAE, a width/height/length and the start image (plus a CLIP Vision output from the clip_vision_h encoder) and return conditioning with the image baked in together with a LATENT of the right shape. Text-to-video uses EmptyHunyuanLatentVideo as the latent instead.

The rest is a normal graph: Load Diffusion Model with wan2.1_i2v_480p_14B_fp8 (or the 2.2 high/low noise pair), Load CLIP with umt5_xxl_fp8 (type wan), Load VAE with wan_2.1_vae, CLIP Text Encode, KSampler (uni_pc or euler, 20-30 steps, cfg 5-6, or cfg 1 with a LightX2V/CausVid LoRA at 4-8 steps), VAE Decode and Video Combine at 16 fps. ModelSamplingSD3 with shift 5-8 sits between the model and the sampler.

Inputs

Name	Type	What it is
positive / negative	CONDITIONING	Prompt conditioning from CLIP Text Encode (umt5).
vae	VAE	wan_2.1_vae.safetensors.
clip_vision_output	CLIP_VISION_OUTPUT	From CLIP Vision Encode with clip_vision_h (I2V 2.1).
start_image	IMAGE	The first frame.
width / height / length	INT	Resolution (832x480 or 1280x720) and frames (81 = 5 s at 16 fps; 4n+1).
batch_size	INT	Usually 1.

Outputs

Name	Type	What it is
positive / negative	CONDITIONING	Conditioning with the image reference.
latent	LATENT	Empty video latent of the requested size for KSampler.

Skip the setup: ComfyUI workflows preinstalled on the cloud

BitVector runs ready-made ComfyUI workflows on its own GPUs. No install, no missing nodes, no red boxes. Open it on your phone or laptop and generate in a minute.

Try BitVector

How to use WanImageToVideo and Wan 2.x video nodes

Download the Wan 2.1 I2V 14B fp8 diffusion model, umt5_xxl_fp8 text encoder, clip_vision_h and the Wan 2.1 VAE into their folders.
Build loaders: Load Diffusion Model, Load CLIP (type wan), Load VAE, Load CLIP Vision.
Load Image -> CLIP Vision Encode; connect it, the prompts, the VAE and the image to WanImageToVideo (832x480, length 81).
ModelSamplingSD3 shift 8 -> KSampler uni_pc, 20 steps, cfg 5, denoise 1.0.
VAE Decode -> Video Combine at 16 fps, h264.

Settings and tips

length must be 4n+1 (17, 33, 49, 65, 81).
LightX2V / CausVid distillation LoRAs allow 4-6 steps at cfg 1: ten times faster.
Wan 2.2 uses two models (high noise, low noise) with two KSampler (Advanced) nodes split by steps; the 5B TI2V model is a single file for 8-12 GB cards.
The 480p model works on 12 GB with fp8 and --lowvram; 720p wants 16-24 GB.
Negative prompt matters: the official long Chinese negative (blur, static, distortion) improves motion.

Set type to wan (and use the fp8 umt5_xxl file to save memory).

Run this workflow from your phone

Every workflow on this page is preinstalled on BitVector with the models and custom nodes already in place. Pick one, type a prompt, done. Zero setup, nothing to download.

Try BitVector

Questions about WanImageToVideo and Wan 2.x video nodes

Which Wan model should I pick?

Wan 2.2 14B I2V (high+low noise, fp8) is the best quality; Wan 2.2 5B TI2V fits small GPUs; Wan 2.1 14B I2V 480p is the stable choice with the most LoRAs.

Can I control the camera or motion?

Yes: Wan Fun Control (depth/pose video), VACE for references and inpainting, and motion LoRAs from the catalog.

Text-to-video?

Use EmptyHunyuanLatentVideo as the latent with the Wan T2V model; the rest of the graph is the same.

Related nodes

Load Diffusion Model

ComfyUI core

Load Diffusion Model (UNETLoader) loads a bare denoising network such as Flux, SD3.5, Wan or HunyuanVideo from models/diffusion_models, with a weight_dtype option for fp8 to save VRAM.

Load VAE

ComfyUI core

Load VAE (VAELoader) loads a standalone VAE file from models/vae so a checkpoint with a missing or weak VAE decodes clean colours, and so split models like Flux and Wan get their decoder.

Video Combine (VideoHelperSuite)

ComfyUI-VideoHelperSuite

Video Combine turns an IMAGE batch of frames into an MP4, WebM, GIF or image sequence at a chosen frame rate, with optional audio, loop count and ping-pong, the standard output node for every video workflow.

KSampler (Advanced)

ComfyUI core

KSampler (Advanced) is KSampler with step-range control: add_noise, start_at_step, end_at_step and return_with_leftover_noise, used for base + refiner chains and two-model sampling.

EmptySD3LatentImage

ComfyUI core

EmptySD3LatentImage creates the 16-channel empty latent that Flux, SD3 and SD3.5 need; it replaces Empty Latent Image in those workflows.

Tired of fixing nodes? Let the cloud do it

BitVector keeps hundreds of ComfyUI workflows installed, updated and tested on fast cloud GPUs. No Python, no CUDA errors, no VRAM limits. Works from any browser.

Try BitVector