<My first lora on here.>
Works well with I2V or T2V.
I trained on the movie clips (36 clips trained at 768 res from cropped 4K footage) as well as still frames and some AI images for the...ones that weren't in the films.
Say: "wearing a blue police uniform"
"with floppy ears" will help fix that if you get 4 ears or if you prompt her wearing a hat.
"wearing a blue hat with orange vest"
*rank 64. 7000 steps
Wan2GP was used for all examples generated with Distilled LTX2.3.
Description
6000 steps total trained using AI Toolkit.
works well for T2V or I2V.
*I'm testing a new workflow to see if i can fix some issues i overlooked.
Sometimes the 0.5 version might have better mouth movement and speech, so try it if some results don't work for you. Less detailed fur but not without it's promise.
same as before, use 'Judy Hopps'.
says 'nude',
or 'wearing a blue police uniform' with or without a hat
or 'a blue hat and an orange vest'
say 'with floppy ears' to get that
FAQ
Comments (26)
Legendary
Are your examples all I2V? I can't get this to work at all for T2V.
Hmm I see that lol. Are you using strength of 1.0? It looks like it might be too low on your generations? Trying increasing it higher? It doesn't do that for me on wan2gp. Are the results with the 0.5 version doing that or just 0.9?
What happens if you replicate one of my exact T2V prompts I used?, try the Money Plane one.
Did you use Judy Hopps with capital first letters? See if adding JudyH@pps does anything. I never saw anything that weird other than 1/100 makes a 2D cartoon style.
I'm liking some prompts more with 0.5 version with the lip sync and sometimes the voice.
Any advice? 0.5 was done to 3000 steps on high noise and 0.9 was to 6000 on balanced. I kept rank 64 throughout. The detail I find better in 0.9, but sometimes the mouth can look not as good as medium distance. I trained only highest resolution and 768 res for the videos clips.
Those samples are with the 0.9 version. I tried with and without "JudyH@pps". Same results. Lora strength of 1.
Used your Money Plane prompt... better but it doesn't look like Judy at all. I will post it.
@playtime_ai_ ok I see. that looks EXACTLY like a test I did WITHOUT my lora activated. Very strange, try that same prompt without the lora and see if it looks the exact same. I bet it does. I can't explain that. What does it look like with the 0.5 version of my lora?
EDIT:
ok even weirder, JudyH@pps only works sometimes, when it doesn't it's just a human. I might try training again using Judy Hopps instead to see if it will fix this behaviour. Before my previous lora attempt got me half the time a 2D cartoon output, so I thought using a unique word might fix that.
@playtime_ai_ can you try using JudyH@pps and see? I actually tried it and it does work with my 0.9 version but it didn't with my 0.5. so maybe it just needed more time training with that in the caption, however, Judy Hopps still works for me.
@kronos1959777 I already did and it didn't change anything.
I'm going to try a different workflow just to take a possible workflow error out of the equation.
@playtime_ai_ @playtime_ai_ just curious if you have any advice about the caption Judy Hopps. Clearly LTX already has an idea that Judy Hopps is a CGI character animal of some sort. If I run this again (i fixed errored captions and too long clips I didn't notice), would you think Judy Hopps of JudyH@pps as the name of the noun for the character is best?
@kronos1959777 When I trained kermit the frog, I captioned him as kermit_the_frog. It seems to help... but it works if you say kermit the frog without the underscores. My cameltoe model works without using the word cameltoe at all. I don't think trigger words really do much with LTX.
So, I downloaded one of my example images from one of my models and loaded it into comfyui to sort of roll back my workflow to a previous save... and it worked! The problem was my workflow. Somehow I messed it up. Sorry for wasting your time!
@playtime_ai_ no problem. Glad it works. Thanks for the new example. Those were...interesting failures I'm sure some people still saved lol
EDIT(oh weird, must be luck of the draw. just got one myself on my new lora test using sigmoid. like 1/100, slightly human proportions and nearly photo realistic environment)
It's not worth investigating more, and I will try to replicate your prompt and see later, but I suspect something is still off.
Yours looks a bit too human in the proportions, and I have never seen it have anything but a CGI background unless I lowered the strength or had another lora activated as well.
You can see the human character in my Avengers video because I used the cinematic hard cut lora
Download it quickly before it gets deleted.
any future plans for Porsha from Sing 2?
Lol nah. Someone might. Only top tier for me. 😜
Lol just googled it. Definitely no. I have a dog looks just like that 😅 give it a shot. Ai toolkit is very good. Even image only training for ltx2.3 is good
How dare you people force me to download this and make a bunch of videos of her eating large pickles without using her teeth and only her tongue and lips to swallow it whole.
You tricked me!
Pickles? I trained this to make her eat carrots...and nothing else 😈
Ay great work! I was gonna try to put her into Wan, but this is even better, lol. :D
I am amazed by the results. you sir have made a gem
Well made, are you planning any more like these or are you open to suggestions?
Gonna do more characters. I uploaded a new version of this one, but civitai is buggy and it cant be viewed even by me through search. Trying to perfect workflow.
@kronos1959777 to find your model I have to go directly to your profile or else it shows up in search but 404 when clicked