SmoothMix Animations LTX

"Epic stuff or lewd stuff... make your choice!"

Its still not on the same level as the WAN 2.2 version. LTX is very stubborn. T_T

So its still a WIP!

The main problem is the inconsistency of the results and LTX has many little configurations, both in prompting and setup, that can really affect the results. Testing it properly is kinda of a pain. =/

But its still possible to make alot of fun stuff.

There is a lot to explain so lets get started!

• ComfyUI Workflow Settings

All videos on all Showcases used this setup:

->Workflow: Smooth Workflow LTX (I2V/T2V) on its default settings (Check the workflow Showcase for more examples);

->Model: ltx-2.3-22b-dev_transformer_only_fp8_scaled;

->Lora: ltx-2.3-22b-distilled-lora-dynamic_fro09_avg_rank_105_bf16 set to 0.5

With SmoothMix LTX set with 0.8 or 1.0 weight.

• Trigger Words

The Loras use the following trigger words:

'SmoothMixAnime' and 'Anime Style' - Will make the video have the style of SmoothMix Illustrious/NoobAI/Illustrious2+NoobAI, I recommend using both;

'SmoothMixRealism' and 'Realistic Style' - Will make the video have the style of SmoothMix Realism, I recommend using both.

• Img2Vid and Txt2Vid

Its primary usage is for T2V. LTX 2.3 seems to have a tendency towards realism, causing some facial distortions on semi-realism/anime images on animations. Try using this Lora with the 'SmoothMixAnime' with weight of at least 0.5 - it might help aliviate that. For T2V I recommend using at least 0.8 weight for most cases.

• Limitations

So far, LTX is proving to be very annoying for trainning SmoothMix Animation. Using the same dataset with same prompts did not work at all - so making a completely new dataset of videos was necessary. After a few tests this Lora was the best result so far - but its still not as 'energetic' as the Lora for WAN 2.2. The main issue so far is the inconsistency:

1. Bouncing and movement: The bouncing is nowhere near as it should be (NOOOOOO!!!! T_T). You can still prompt it and get a great results but it doesn't happen everytime. Movement can still be very stiff.

2. Uncensoring: The uncensoring - overall - works but sometimes can make some weird stuff. (NOOOOOO!!! T_T). So its still not where I want it but it should work well most of the times. I added the male dataset used on Futanaris and Males Lora for WAN 2.2 to uncensor the guys as well but it can only create... huh... eldritch tentacle abominations on the guys (NOOOO!!! T_T), so don't even try those... unless you want to make your eyes bleed! I'm trying to avoid needing to train another Lora just for males and futanaris again but it seems it's going to be inevitable. sigh

As always feedback is appreciated! I'll be working on this specify project for now. My main goal is improving the dataset for now. So if you have any difficulties on making certain poses/actions be sure to tell me as I'll be taking notes! ^^

Actually I'm still not sure if I just need more videos on the dataset or if the training needs more steps... or both. This Lora is already using more videos than the WAN 2.2 Lora - with higher quality - and yet the results can be very different between the two.

The training takes WAY LONGER as well - 30 hours to make this! WTF!? I'm sorry WAN 2.2! Perhaps I treated you harshly! T_T

Anyway, this is only possible with your support! =) If you wish to support a bit just visit my DeviantArt page. Any support counts. ^^

Thank you for all your support!

Thats all for now!

Have fun! =)