Update 8/14/2025
V2 introduces a custom node I built called EA Motion Bias (Lightning) which gives you presets for controlling the motion bias curve. The image showcase is a sneak peak of my next series, before post processing. It is shaping up to look smooth. I've bumped up to a 5 step 3/2 high/low split to make motion more aggressive. The node tunes the sampler scheduler and other parameters as well as lightning strengths to give more flexibility and encourage experimentation.
Recommend changing the preset drop down and setting WanVideo sampler scheduler and high / low cfg based on what it prints out at the bottom of the node (it lists all the parameters its setting). All the other sampler and lightning settings (steps, sigmas, noise, etc.) are auto-tuned based on the preset.
General
This is a reworked version of the WAN 2.2 I2V 14B lightx2v Lightning workflow here: https://huggingface.co/lightx2v/Wan2.2-Lightning/tree/main/Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1
This has been tuned to cap out at about 23.1GB VRAM (to fit on a 4090). It uses WAN 2.2 fp8_scaled high / low models and VAE tiling. I generate a 6 second 16fps video in 120-130 seconds (once T5 is warmed up). The videos are high quality and do not suffer from snapback. I do all my videos in vertical 16:9 for phone viewing so this has been tuned for that.
I tried a few of the other workflows that have been uploaded but they had a bunch of custom nodes that did not install well for me so I've kept this one dead simple, you just need to download the scaled models and the high / low noise lightning LoRAs (links in notes).
Description
This version introduces a custom node I built called EA Motion Bias (Lightning). This node has built-in presets for controlling the motion / detail curve by tuning the high / low samplers sigmas, cfg, scheduler, steps / step split, add noise, and lightning lora strengths.
This lets you quickly jump between presets to get different effects. The default lightning is included in the preset list as Lightning Default 2/2 (should give you the same output as v1 of this workflow). On my latest generations, I've been using an Aggressive 3/2 split which gives an extra step to the motion. Each of these presets ensure that all settings remained balanced for the scheduler they use.
FAQ
Comments (5)
Thanks for this. New update with Motion Bias fixed some of the loras not applying their effects fully. Do you know if there is a general lora strength rule? Like same str with high&low or half str on the low? I'm going with 1 str high & low on each lora atm. Also I only set Lightning to use Lora Select node and connecting other loras to it with Lora Select Multi. Is there any downside to this?
1 is good default, as low increases it can slow down the motion (same with low steps). There is a balance between all these variables and it varies with scheduler / sigmas (all driven by motion bias in v2 workflow). The aggressive preset is tuned for maximum motion (3 high steps, 2 low steps) and that's what I'm currently using.
Probably no downside, just needs to have the lora in / out (I will probably add that node to a future version to save space). All loras are weighted deltas so order is not important.
Working on a custom post processing / pre-training multi-purpose workflow with a set of custom nodes right now but will make more updates here when there are useful additions or better tunes.
WanVideoImageToVideoEncode
shape '[1, 320, 1, 2, 26, 2, 15, 2]' is invalid for input of size 1017600
Maybe some BF16/fp16 error in the model settings+
Eee.A....Sports.....It's in the game
there is a node not found in comfyui