Trained on some low quality 2000s era images to push LTX into generating some more grounded t2v outputs easier without heavy prompting. Removes the LTX high contrast gloss look pretty well.
PROS :
- Ground the LTX look a bit, reduces contrast. Introduces a touch more amateur haze and blooms.
- A little bit of nudity capability bonus, but this wasn't the focus here.
CONS:
- tends to whitewash people. Dataset wasn't diverse enough. Will try and source a better dataset for next training round.
- Sometimes adds booba to men (maybe this is your jam?) but it wasn't intentional.
- Slight motion reduction.
Description
FAQ
Comments (28)
Looks like fun :)
Looks promising. Very cool. Nice work!
Dude this looks amazing, and only rank 16?
Is that better for this type of lora?
Looks like basically upgrading it to LTX 2.4!
How many images trained roughly? Resolution used?
I trained in 64 to make sure I catch everything, but then downconverted to 16 afterwards. I have a 32 conversion as well but 16 is plenty.
it's just a style lora, trained on about 500 images of lower quality 2000s era snapshots. It adopted the style pretty early in training.
res- I don't recall. I think it was pretty low at 720 longest edge I think.
What program did you use for training?
@bionovafood863 my own interface but the backend is Musubi
This looks good. My first few tests went very well. What is your go-to weight for the Lora?
I just throw it on full power: 1.0 . worked fine for me that way.
The only problem might come from some gens with tight white shirts as you might get some nipples poking through since there is some topless content in the dataset. but it's not a huge issue.
unwanted nipples is the only time I'd drop the strength
Very nice! Have you tried it on I2V yet? Or only T2V?
it would def work on i2v but there's no real point as when you use i2v you're setting the look via your start image, so the lora can't really do much for you there. . unless you want your i2v to show a boob - then sure. this will help with that.
@QualityControl seems like there is a lot of point, since i2v is ass in ltx, maybe a high quality lora will help it out somehow
@brendacaro let me know how it goes. I haven't really tried it, but hey, maybe it's accidentally great
nice very nice
Good work, It work on ITV too. It make the action a lot better and more follow the prompt. As it somehow will change the image a bit, turn strength down a bit will have better result.
Great Lora. That's how orginal Ltx should looks. Many thanks
Funny, the link is down, maybe is a civitai issue, but you should check it out
prob a civit teething issue with this new two front doors thing. This is a SFW lora so not sure where it got pushed to. might have gotten mixed up somewhere. It'll prob get resolved soon. it's fine on my side
@QualityControl it has been resolved
@FlowSpecial good to know. I did nothing :) figured it was a migration thing
Loving this Lora! I haven't really ran into the whitewashing yet, definitely had the male booba issue lol. I think there's a ton of potential here, Excited to see future iterations!
Whats the recommended lora strength?
0.7 - 1.0 . all my example gens are made at 1.0
its only for T2V cuz didnt work in I2V or is there some trigger word ?
it's mostly only useful for t2v because it's a thing that sets the look of things. with i2v you're already establishing the look with your input image, thus it wont be noticeable.
no trigger word needed.
my image is good but ltx skin is fuckd up what did u use for skin or its just like that as default ?
@hotsizzle1 i didnt use anything extra. just typocal ltx workflow and my lora at 0.8-1.0 . any other issues are likely from ltx itself
It's amazing. Honestly, it doesn't feel like an addition. It's like I'm using a completely new level of modeling. This lora is a revelation to me. The only downside is that it lacks naturalness and realism, but I hope this can be addressed with other loras. 100/10
Fantastic. Works great with my realistic style. It makes the characters look like actual people you might meet and know.
Details
Files
AmateurHour_01_rank16.safetensors
Mirrors
AmateurHour_01_rank16.safetensors
AmateurHour_01_rank16.safetensors
AmateurHour_01_rank16.safetensors
AmateurHour_01_rank16.safetensors
AmateurHour_01_rank16.safetensors
AmateurHour_01_rank16.safetensors
AmateurHour_01_rank16.safetensors
AmateurHour_01_rank16.safetensors
AmateurHour_01_rank16.safetensors
AmateurHour_01_rank16.safetensors
AmateurHour_01_rank16.safetensors
Amateur Hour - LTX2.3.safetensors
AmateurHour_01_rank16-mid_2530917-vid_2844417.safetensors
Amateur Hour - LTX 2.3_AmateurHour_01_rank16.safetensors
Amateur Hour - LTX2.3.safetensors
amateur-hour-ltx-2.3.safetensors
Amateur Hour - LTX2.3.safetensors
AmateurHour_01_rank16-mid_2530917-vid_2844417.safetensors
AmateurHour_01_rank16.safetensors
AmateurHour_01_rank16.safetensors
Amateur Hour rank16 - LTX 2.3.safetensors
Amateur Hour - LTX2.3.safetensors
Amateur Hour - LTX2.3.safetensors
Amateur Hour rank16 - LTX 2.3.safetensors