Home
Creators
Search Recent Random
Posts
Search DMs Popular Hash Lookup Tags Random
Importer
Import FAQ
Register Favorites Keys Review DMs Logout
Community
Telegram ThePornDude
Home Creators Posts Import Register Favorites Logout

WAN 2.2 - Text, Image and Video to Video (Patreon)

Published:
2025-08-05 18:47:55
Edited:
2025-11-20 20:17:55
Imported:
2026-09

Downloads

Content

Video - https://youtu.be/MkAuWTLJp1s

Overview

WAN 2.2 is fun... but there are quite a lot of options! In summary they are:

5B - Image & Text to Video all-in-one

14B - Two sets of two models:

Text2Video - High & Low

Image2Video - High and Low

Wan 2.2 - 5B

The 5B model makes great quality video, but isn't as good at following the prompt as 14B. Speeds are fairly similar as 14B can be accelerated to just 6 steps, making each around 6 minutes on a 3090.

5B uses a lot less VRAM too, so for 8-16GB cards this may be your best option in the speed vs performance game.

5B is also a 720p model and lower resolutions can be a bit pants. I made the mistake of starting out at 480p and wondering why things looked bad :)

30 steps, dpm++_sde, cfg 3, shift 9 is typically decent but play around and see what you prefer. High cfg values tend to "burn" and lower shift values may not follow the prompt as well.

ComfyUI repackaged model - https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/main/split_files/diffusion_models/wan2.2_ti2v_5B_fp16.safetensors

WAN 2.2 - 14B

The king of the jungle, but also a very hungry beast. FP8 quantisation helps on a 3090, and the same LoRAs as last week also give 2.2 a much needed speed boost by reducing the step requirement to 6.

Things that caught me out

Resolutions seemed a bit strange at first, in that 1280x720 gave errors but 1280x704 was OK. Not sure if that is how they're to work, but it works for now!

Workflows

Both 5B and 14B workflows are attached. The defaults should be good to go, but experiment away and do let me know what things you find!

They follow on from last week's WAN 2.1 video, so it may be worth starting there if you haven't already - https://www.patreon.com/posts/wan-t2v-i2v-and-135155677

Links to the various models you'll need are in the workflow, so you can just click to download.

Most files you can get from:

https://huggingface.co/Kijai/WanVideo_comfy/tree/main

https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled

As with the WAN 2.1 nodes, you'll also need the various custom nodes installed, as detailed in the workflow notes:

4 Custom Nodes - Install using ComfyUI Manager:

ComfyUI Wan Video Wrapper - https://github.com/kijai/ComfyUI-WanVideoWrapper

KJ Nodes - https://github.com/kijai/ComfyUI-KJNodes

Video Helper Suite - https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite

RGthree - https://github.com/rgthree/rgthree-comfy

Enjoy! :)

Files

Comments

Wicked Desires

FYI, the settings (as loaded from the json file) for image 2 video in the WAN_2.2_14B_Kijai_v13.json do not work. "Merged" from lora needs to be set to false, and the start and end frames for the samplers needed to be adjusted.

Davinzzie

After a 1st succesful generation of video2video I get the error: index 0 is out of bounds for dimension 0 with size 0

Scopitones

New here. Thanks for sharing this. When trying to run text2video I seem to be getting this error "Error(s) in loading state_dict for WanVideoVAE: Missing key(s) in state_dict: "model.encoder.downsamples.0.residual.0.gamma", "model.encoder.downsamples.0.residual.2.weight", "model.encoder.downsamples.0.residual.2.bias", It goes on a lot longer but I dont want to spam here. Any help would be appreciated

Carlos L. Romero

getting this error, any idea? Sizes of tensors must match except in dimension 0. Expected size 160 but got size 80 for tensor number 1 in the list. this was trying to do a i2v

Han Solo

Hi NerdyRodent! Thanks for the hard testing work you put in those workflows :) I would like to suggest a workflow idea that I had. I tested it in Pinokio>Wan's GUI and it works amazingly well. With an FLF2V workflow you'd be offering the best of I2V and FLF2V in one workflow: while I2V is powerful, an FLF2V setup offers all the same capabilities plus some unique advantages. WanFirstLastFrameToVideo node includes inputs for both a start_Image and an end_Image. By bypassing one of these inputs, and using a SINGLE image as input, you can achieve more than a standard I2V: - Generate the future: provide start_image, video will generate forward from that frame (standard I2V). - Generate the past: provide end_image, video will generate the events that led up to that image. - because both generated videos use the same input image, past and future will flow one way. No more issue of rain falling upwards or any other physics inversed, because one decided to play a generated I2V result in reverse in order to make a video feel longer. With WanFirstLastFrameToVideo node you'd be offering the best of both I2V / FLF2V worlds in a single .json Thanks for considering my request ;)

Gregory Plichta

I'm probably asking a stupid question but I am getting an Error "No input found for flattened id [169] slot [-1}. I'm trying to do the Video2Video Option on the 14B

Gregory Plichta

Ok I fixed my first issue but now I’m getting the error about the SageAttention. Your settings for the WanVideoWrapper are the same as mine for attention_mode set to sageattn. Am I missing something?

colin-vt

I was curious if you have thoughts on how to improve the speed of the generation? when I use the wan2.2 example workflow i can generate a video in around 70-90 seconds, but this flow takes over ten minutes (running I2V on a 4090). Is it just because of the output size being 720?

Light Scheme

Hey @NerdyRodent would you know why the 5b video to video workflow would be outputting very odd bleached blue results with no definition or resemblance to the original video input?