lightx2v

/

Wan2.2-Lightning

like 440

On Sep 28, 2025, we released two models, Wan2.2-T2V-A14B-4steps-lora-250928 and Wan2.2-T2V-A14B-4steps-250928-dyno. The two models share the same low-noise weight. Wan2.2-T2V-A14B-4steps-250928-dyno delivers superior motion rendering and camera response, with object movement speeds that closely match those of the base model. For projects requiring highly dynamic visuals, we strongly recommend using Wan2.2-T2V-A14B-4steps-250928-dyno. Below, you will find some showcases for reference.

It is important to note that Wan2.2-T2V-A14B-4steps-250928-dyno's high-noise model was trained using a full-parameter strategy, which has necessitated some adjustments to the workflow. The updated workflow has been detailed here. Please also be aware that this model may occasionally produce visual artifacts—such as blurring, ghosting and geometry intersection—particularly at 720p resolution.

Both models have their limitations and significant room for improvement. Nevertheless, they represent our current progress in enhancing motion speed, and we hope they may still offer value to the community. We are committed to continuous iteration and welcome any feedback on issues encountered while using these versions.

Thank you so much for your work guys!

this is amazing!!! soo excited to try!

I'm not sure why, but I seem to get far worse results with these than with the previous versions - both for the official and Kijai 2.2 Lightning Loras...
What is the suggested High and Low Noise weight for these? Should both be set to 1 with CFG also at 1?
I tried loading up the workflows that are here in this repo, but for some reason, they won't load in Comfy for me and I also can't drag and drop them in...

So you're telling me I've waited 3 months for a T2V improvement? Bummer. What about I2V?

Waiting for dyno low noise version (full model), thanks for the work!

Need help :( I use just standard T2V-A14B-HighNoise-Q4_K_S.gguf and Wan2.2-T2V-A14B-LowNoise-Q4_K_S.gguf, and after connecting 250928 lora, both high and low, i get some weird quailty in first few frames. Like there is a conflict between high and low in 250928 version. Or is it not meant to be used with standard wan2.2 t2v? What could be wrong?
Update: Now it somehow works... The problem was video length it seems. Because of my low end system, i sometimes set the length to very low like 9 or 17, about 1 sek of video. After i tried 49 (3sek) 640x640, that worked. And maybe lowvram mode plays a role too...

I can't get these Wan 2.2 Lightning Loras to work at all - neither these official ones or Kijai's versions.
At the moment, the best results I'm getting are with Wan 2.1 Lightx2v loras, set around 3 strength on the high sampler and 1.5 on the low. Not perfect, but reduced slow motion and better general adherence. The new ones are working even worse for me than the previous Wan 2.2 versions.... 🤷‍♂️

outputs with new loras have more detail but motion is still much worse than wan 2.1 light loras

the Dyno is actually the Wan 2.2 T2V model itself?

i don't understand anything, when i put the model ( it seems to be a wan model itself , if i drop the exemples above in my comfyui worflow) the diffusion loader, animation is not better, but when i load it like a lora, quality is like a Q8 with 20 steps very good, i didn't even know i can load a model in the lora loader, and that works .. better
i explain:
High model Wan2.2-T2V-A14B-4steps-250928-dyno
Low model Wan2.2-T2V Q5 GGUF + lightx lora

Result a standard quality

But

High Wan2.2-T2V Q5 + Wan2.2-T2V-A14B-4steps-250928-dyno (in the lora loader) 2 steps cfg1
Low Wan2.2-T2V Q5 + lightx2v rank 128 2 steps cfg1

Excellent prompt adherence

Edit: I'm just dumbass, of course this can't be load as a lora, i have so much error in my log , it's seems that 4 steps without any speed up lora even with a cfg at 1 give me a good result because of Wan Moe Ksampler that understand that at 4 steps it needs 3 steps in low, the all thing was just done b y the low model so..