LTX2.3 FP4 Models

yes, but other people are uploading those, I just started sharing the fp4's because no one shares fp4's on civit

lol, yeah, my hard drive is mad at me, not only do i have all these different fp4's, i have the full bf16's for working on my custom fp4's.. 250gb just in ltx2.3 models.

@artificialotaku yeah, add in all the other stuff like hunyuan video, ltx 2, flux2, etc.. yeah, this 2tb nvme is just for comfy...

i never used "default" ltx workflows, they never work for me because multigpu.
i use vaeloader kj audio for the audio vae, vaeloader multigpu for video vae, and dualcliploader for text projection + gemma 3.
i don't even think I have seen the default workflows for ltx2.3, i just learned from the ltx2 ones they don't work at all for me.

oh and updating comfyui, not sure what version you are using, but i know on the portable comfy, the update inside the program itself doesn't work right on these major updates, you need to use the updates inside the update folder, there is an option for update everything that will update comfy and all python crap, but, be careful, a lot of shit broke in that update.

@Foxdude yeah, i haven't use the desktop app yet, but, have been working on building my workflows for now on with the "app" thing configured. i like the app thing really, kind of nice when im laying in bed playing with it on my phone.

@Foxdude lol, that is EXACTLY why i got into ai video, i make a visual novel, 3d render everything, animations were always day long processes, but, with ai video, i can just render key frames now and let ai interpolate.

@Foxdude biggest problem was though, i got everything setup and ready to go in ltx2. then 2.3 dropped. like really?! damnit, start all over.

@LetTheBassDrop yeah, comfy is a huge mess atm, i refused to update for the longest time, but kind of had to for ltx2.3, 16 hours to get it running again, and i still haven't fixed everything, just what i needed to get it running. and if i remove --disable-dynamic-vram from my batch file, it won't even start, just crash with no message. the dynamic vram should of been optional, not forced on everyone, it broke sooo sooo much.

you don't need the distilled lora on the distilled model, you don't need the distilled on full either if you want to run full steps.

@BigSad11 ok, so im not the only one, lol, i tried and never got ltx2 or 2.3 to work without distilled lora or the distilled model, i just thought it was me.
i never got the base model work in either ltx2 or 2.3 without distilled, fp4 or fp8, so unsure, not sure what I am missing, i tried 25 steps, 50 steps, 75 steps, 100 steps, 200 steps.. never got a clean output without distilled lora, and honestly i found running distilled lora at low strength gave great results at 25 steps, but highest step count i get get results from was like 32 steps, no matter the strength of distilled or with it off, i could never get results over 32 steps.

@BigSad11 set distilled lora to about 65%, thats kind of where i found the happy point (in ltx2 anyways), was good at 8-10 steps. if you want to push to higher steps and let base model come out more, try about 30%, thats where I was hitting 24-30 steps.
and yeah, distilled inherently snaps the model into a quick output, so less likely to follow prompt that well.
what i tend to do, is kind of "undistill" with lora's, i will inject a bunch of lora's at low strength, it returns some of the creativity lost by the distillation.
if im doing general outputs ill inject a bunch of things like vhs, galaxyace, etc, any sfw loras between 5 and 15% (the more lora's, the less strength)
and of course, for nsfw, yeah, a bunch of nsfw stuff (except pov ones, those snap hard even at low strength).
this just injects creativity back into the distilled model allowing you to raise step count a bit and prompt better.

@BigSad11 yeah, its been a bumpy road with this model, and here in a few days, may not matter as much anymore. Magihuman is out, its pretty much wan with audio, so knowing how much everyone loves wan, lol

im waiting in kijai to finish the comfy wrapper for it, then ill be making fp4's, its a smaller model so should be around 8-10gb if i can maintain same level of quantification.

i probably won't use magihuman as much since it is basicly just wan with audio, wan has that 8 second limit and doesn't look like magihuman changed that, so ill stick with ltx2.3, i like my 30-60 second generations.

one weird thing im investigating with magihuman, is how i watched it generate a 5 second clip, in 3 minutes, on cpu only..

@BigSad11 yeah, magi human just dropped like 4-5 days ago, and its architecture from what I have read, is basically wan and stable audio glued together.. its more complicated then that, but that's the easy way to put it, im just waiting for kijai to finish the wrapper before i dig into it more, ill be curious to see if wan lora's work on it.

and yeah, i had a lot of hopes for hunyuan, i liked it better then wan, then hunyuan 1.5 dropped, it was great, i was training lora's for it to add all the adult stuff for people, then ltx2 dropped like the next month, and that killed hyv1.5, its a good video model, but now we have opensource video+audio, video only was kind of dead after that.
crazy thing is, my fp4 of hvy1.5 is 5gb.
i still feel that hyv1.5 has superior video, but that lack of audio now really hurt it, lets hope that tencent is working hard to add audio to hyv2.

comfy up to date? and i personally have to load this with a checkpoint loader instead of unet loader because of the same issue.