i2v 0.5 - Helps to reliably grab neck
Sometimes adds too much , like extra hands or grabbing himself/herself...
Should work with multiple angles , search for right seed and balance of lora strength.
Merged test with doggy/missionary lora to be less static. Audio might unsync.
Posted as is , don't expect another ver
Description
FAQ
Comments (7)
Great work!! High quality!
your loras are as good as wan2.2 whats the secret? learning rate? can you share your training args? ive just being wasting time with ltx2 training with settings that wont work as good...
For the sake of the LTX community. Please answer this @blo01! Your Loras are amazing.
Thanks,
I think it's just i2v training at 1.0 first frame cond , tried t2v in 2.0 and that didn't turn out that good . Maybe 2.3 is better , didn' try yet.
Sampling seed matters a lot , these still turn into body horror as grab lora was still undercooked.
Other than that you just hammer lora with same concept.
20-30 videos seems like a sweet spot where it's diverse enough and doesn't take ages to converge. 2.3 seems to take longer than 2.0 but it's generally better so it cancels out.
For parameters just default ones are okay :
rank - 64 seems better , 32 is hit or miss
optimizer - AdamW/8bit , still trying few but have doubts they do better , this one at least has down trend
lr - 0.00005-0.0001
res - 768 , 512 also works but a bit higher works better ? Could be placebo but I think 768 converges faster
frames - 97 , which is 4 seconds , 121 just takes longer/uses more vram , can even slice to half and train on 49+97 if it's repetitive action
fps - 25
You can train video first to get it right, then add audio as well by training on top with both datasets. This one is a bit experimental merge trying to add 2 concepts so audio unsyncs sometimes I think.
Still takes around 3k steps and maybe around 5k it gets the concept polished
Also 2.3 maybe is better at erasing dicks/nipples not sure.
Also saw that big uncensored lora training has started ,maybe better to just wait and build on top of it even if's t2v still should help.
@blo01 thanks for sharing.
i trained my self anal fisting but it overfitted really fast, bleeding the videos all over ignoring prompts.. that dataset is also small and only 2 or 3 different scenes, not plenty like my other loras.. i will try your settings and others i had pending this time with a 3d style videos of missionary sex... wish me luck, i prolly publish m4crom4sti4 one i did that really doesnt look bad besides producing black nipples sometimes in the upscaling phase..
@blo01 Would you consider putting together an LTX2.3 training guide? I'd throw buzz on it, others probably too...
@blo01 Whatever magic you found seems to have at least helped to unlock the door to training LTX2.3.