The goal of this Lora is to get more realistic looking images that can be used for social media purposes. (AI influencer grift)
Trained locally on a 5090 on WAN 2.1 14B with 2600 Joycaption captioned social media scraped images for 2800 steps (1400 steps/ batch size 2).
Samples are generated with my custom WAN 2.2 text to image workflow
Left is no Instagooner, right is with instagooner applied
Same seed for every pair
Lora strength should be:
2 on WAN 2.2 HIGH noise
1 on WAN 2.2 LOW noise
All faces were removed, generated faces are not real
Hit me up if you need a WAN based AI Influencer LORA for a consistent face
Description
FAQ
Comments (23)
I don't understand- if you trained for wan2.2, why on Earth does one LoRA need a strength of 2- this seemingly makes no mathematical strength. 1 is default so the maths in the LoRA should surely take account of this.
To me, this suggests current training for wan2.2 is fundamentally wrong in some way. If it were not, the base strength for both high and low LoRA would be 1.
Your lora looks really amazing but could you please tell me if you've used other loras like Lightx2v or enhances for your example pictures? You've used 2.2? If you have a workflow or more indications it would be greatly appreciated 🙏
Thanks, great job. Is there a way to implement Lightx2v for Wan 2.2 lora in your workflow?
just replace the loras in the workflow with the new lightx2v ones....
sukuna126853 I did but the result is bad. there must be another way.
MuhHo as those are brand new I haven't had the chance to test yet will check tho
IceKiub thanks, appreciated.
MuhHo heard the models were bugged and need to be used aat 0.125 strenght instead of 1 maybe that might fix it, kijai also released version that work normally
looks great, what lora rank/learning rate did you use?
Appreciate it
rank 64/alpha16
4e-2
2670 images for 1400 steps at batch size 2
Amazing Lora. Works well on Wan 2.2 (pictures and videos) and it does its magic also on SFW content, using it for everything right now :D
Appreciate it!
Can you use something say from flux (Using a lora) and have this make it more realistic?
no idea you should try tho
Just to confirm. When you say "Left is no Instagooner, right is with instagooner applied" what exactly that means? They only show on the left?
Left is without lora right is with
IceKiub oh..you mean in the picture sample.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.