RuneXX

/

LTX-2.3-Workflows

like 385

This workflow will be updated, will try make it work better, currently it doesnt really allow much changes.

Congratulations!!! This is just amazing.... it even works really well :)

Thanks very much for the workflow. This works perfectly for the audio. I can use any video with any actor/actress and it clones there voice perfectly. Why is this not possible with your Longer Video workflows? The voice carry over never matches and sounds terrible in comparison to this. I don't get it. Obviously this workflow shows it is possible to clone a voice accurately with LTX 2.3. Any ideas why it only works properly with this one? The length doesn't seem to matter here. I can make a 20 second recut for a video with this one and the voice is a perfect match from start to finish.

. Why is this not possible with your Longer Video workflows? The voice carry over never matches

In the Long Video workflow multiple videos are created in a loop, or in groups (depending on what workflow you used).
So in a way its multiple independent video parts created. And in the end, all these videos are glued together as one long video.
It does have some reference input before creating a new video part, about 73 frames (but you can set this longer in the workflow settings at bottom). If there is no audio in those last 73 frames (voice audio), the next video part has no idea what was created before and will create an entirely new voice audio. Its a limitation the way it is currently

So the work-around would be to add custom audio input ... for example a song. As used in the examples.
But it could of course be one long monolog input audio as well

Now, that being said, luckily we have a new lora out, called ID-Lora, that should be able to use 5 second reference audio in every video part in the long video.
By this the voice should be consistent (for one character). Will update a long ID-Lora workflow soon.

LTX can only create video with same voice audio, if it knows what voice audio to use ;-) either by custom audio input, or by having enough reference frames that include voice audio when extending or editing

Thank you very much for the detailed explanation. Your very helpful. Also, thanks again for this recut workflow. I'm having a blast messing around with it.

By the way, this is better then any voice cloning model I've ever used. I've tried many of the recent ones. This surpasses them all by a long shot. In my opinion, there all crap compared to this.

LTX can be quite surprising sometimes.. . Its a bit of a "jack of all trades" .. ;-)

I had a surprised moment when i just input a drum solo audio clip and a static image of a drummer, and it pretty much nailed all the drum hits (kind of, at least good enough for a regular non drummer viewer)
And for voice cloning its surprisingly good. It can even dub to other languages (made a workflow for that for LTX-2.0, will update it to LTX-2.3).

Hi RuneXX,

First of all, thanks for sharing these LTX-2.3 workflows! I've been trying to replace video segments with V2V ReTake and have a technical issue I'd like to discuss: artifacts persisting during regeneration.
I'm working with a source video (also generated with LTX-2.3) that has a specific segment with poor frame joining, camera shake, and artifacts. When I try to replace that specific segment using your workflow, the output consistently shows similar artifacts or inconsistent motion, almost as if the model is "contaminated" by the original bad frames it's supposed to replace. I was careful to select the correct time range that includes these artifacts.

My question is whether this behavior is possible or am I doing something wrong. Is it possible that the new video segment is being influenced by the original underlying pixels?

I also wanted to ask if you've ever created an audio/video extension workflow tied to the final frame and possibly also the median frame.
Thanks for everything!

Will try experiment with something that is similar, see how strong the masking effect is (its rather weak probably, and it does take into account the original frames).
If its too much of a change it might struggle. But I'll try see if i can find a way to push the limits. Maybe even add some dark overlay over the frames to change, to make them have lesser impact.

Will try see if it can be improved for a stronger effect ;-)

Thanks for your reply! I just wanted to mention that V2V ReTake has worked really well in other situations; for example, when I have the same character say different things. Best regards!

Thanks for your reply! I just wanted to mention that V2V ReTake has worked really well in other situations; for example, when I have the same character say different things. Best regards!

Yes for small changes, or "logic" changes all seems well.
But might be possible to make it stronger to be able to completely change things. Will try with some more heavy "masking" out

so, if someone is wearing a blue shirt, can i change the color of the shirt with this WF?

so, if someone is wearing a blue shirt, can i change the color of the shirt with this WF?

Yes in theory. If that person is at an "inner section" of the video (in other words not present at start or end). Since this workflow is not inpainting/masking all of the video but rather re-shooting/re-doing a part of the video. There are other workflows for inpainting that might work better though. Depending on what exactly you want to achieve.

I'll upload a more generic inpaint workflow soon. That is probably more what you might want. Where you can replace, change or edit certain parts of the video.