From versions 1.0 to 4.0, my primary focus was eliminating the "moonwalk" effect that typically appears after 3–5 seconds of walking. After extensive testing, I’ve concluded that this effect is likely an inherent limitation of the Wan 2.2 architecture that cannot be 100% eliminated through training alone.
.
Even after training the model on 20 clips (15 seconds each) covering every possible angle, the improvement was marginal and came at a cost: facial consistency drifted and "glitter" artifacts began to appear.
Seeing similar results in other creators' models reinforced this belief.
.
The Pivot: SVI Integration The introduction of SVI (Spatial-Video Interpolation/Integration) changed the game. I found that by splitting the scene (starting with an 81-frame segment followed by 121-frame parts), the moonwalk issue is handled much more effectively.
With the motion stability solved via workflow, my latest model iterations have shifted focus toward improving breast physics and natural movement instead.
.
Recommended Settings
While newer model versions exist, these specific settings have yielded the most stable walking results for me:
SVI Configuration: 1st Part (81 frames) | Subsequent Parts (81-121 frames)
Checkpoint (High):
wan2.2_i2v_high_noise_14B_Q8_0.ggufCheckpoint (Low):
wan2.2_i2v_low_noise_14B_Q8_0.ggufLoRA:
Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64.safetensorsWeight (High): 3.0
Weight (Low): 1.5
Steps: 6
LoRA Strength: 0.5 – 1.0 (I personally find 0.5 to be the sweet spot)
(Tip for Slow Motion: If you encounter "slow-mo" issues, if you using SVI, simply increase the output frame rate.)
(The breasts motion and physic work well with the M4CROM4STI4 model)
https://civarchive.com/models/1852647/m4crom4sti4-huge-natural-breasts-physics-wan22-video-lora-k3nk
.
Prompting Guide
Trigger Word: wlkng Prompt Strategy:
wlkng,
The subject walks casually alongside the viewer, with the camera following the subject’s movement.
Description
V2.0
Train with a doubled dataset to
fix cases where some images make the model stubbornly refuse to walk.
Strength: 0.7 - 1.0
Trigger word (camera from front) :
wlkng,
she is walking towards camera,
Trigger word (camera from behind) :
(strength for behind should be 0.75)
wlkng,
she is walking,
Trigger word (camera from side) :
wlkng,
she is walking,
.
.
NOTE:
You might notice that sometimes the girl’s body gets pulled back to the original position after walking for a while.
This happens because of Wan2.2 itself, and also because most of the dataset clips are only 4–5 seconds long.
But even when I tried training with 25 clips of 12 seconds each, the results improved alot, but the pulling effect didn’t completely disappear.
Solutions to reduce the pull-back effect:
Avoid using fully front-shot images – the background tends to pull her backward.
Use side view or 45° front-side view images – I tried this and it worked well.
(Optional) Try a new seed or a different input image – sometimes this helps smooth out the walking motion.
(but) if you don't care about position movement you just care about girl walking motion and you okay if her walking in place, so you can try this "the girl is walking on treadmill"
FAQ
Comments (18)
Is a CFG of 3.5 also recommended when using the lightx2v or wan 2.2 lightning loras? I'm pretty sure it is advised to use CFG 1 when applying those loras so just wanted to double check :)
I normally use CFG = 2, but when I tried 2.2 I felt that some images were quite stubborn when making the character walk, so I increased the CFG.
So, for your question — feel free to do some A/B testing. I do that as well.
As for the CFG, I suggest increasing it just in case you experience the character being stubborn about walking, or ending up just walking in place.
Are the jiggling breasts part of the lora or are you using another lora? How does it behave with smaller breasts?
Did the training data perchance have a lot of kling ai walking videos?
I’m not satisfied with any of the Kling AI results for ‘walking’
@baihsan I use the prompt "the subject walks casually alongside the viewer, the camera follows the subject moving" then add any extra bits you want after
@Corbe
nice, how about the result
the file for the high noise model is missing
uploading
don't know why but the server super slow right now
@baihsan Just wanted to make sure you were aware.
@playtime_ai thank you
Love this! Keep on the good work.
Thank bro
does this work for t2v as well?
i have never try t2v
so the models aren't optimised for t2v yet
but you can try, because some lora i still use the wan2.1 version too
I hear that using i2V models for T2V is possible ,, if you will hook an Empty Latent to an i2V model/Loras . Honestly I never tried this before, but someone can confirm this trick to us ?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.