RuneXX

/

LTX-2.3-Workflows

like 618

Yes the extended video has its weak points.

• it uses 73 frames as reference frames (video and audio). That is the last 3 seconds of previous video.

Anything not happening in those last 3 seconds, the next video part has no idea about. So if the voice is not part of the last 3 seconds, its likely to change in next part.
You can set the reference frames higher though. 73 is the minimum recommended by LTX. (and i put that in the wf, since many will probably try extending short 5s videos)

The color drift/changes are strongest in 2-pass workflow. Since it first generate at low resolution version and then runs it again through a 2nd pass upscale.
This introduces new details (often better) than the original, but since it differs from the original it might be quite noticeable

• The workflow works best with single pass (the workflow has a single pass / 2pass mode toggle)
• With 2-pass, it tries to blend the frames in an overlap so the changes are a little less noticeable
• With 2-pass there is also a color match toggle at each group that can help color shifts a little bit (only use this with 2-pass, at least in theory single pass shouldn't need it)

That being said, the color match is set really low to be subtle (0.25 if i remember). You can peek inside the subgraph and set the color match stronger (its near the top right of the subgraph wf).

And for any strangeness (that will happen, 3 seconds and the model has to "guess" what to do next... sort of), the best fix is often just changing the seed and prompt.
If the output was not desirable.

That all being said, its the newest workflow, will take a look and see if it can be made to work even better. Maybe some latent guiding nodes or something

Also color matching node is "dumb" so it takes the whole image. In your examples the green is very strong (background). It will make everything a shade of green likely
So it might not be a node for that particular video at least, and best be turned off

A little test run with no color matching turned on, and single pass... seems to hold up pretty well. Ideally this type of workflows should run at single-pass

But its the 2 pass workflow that's probably the challenge, as 2 samplers introduce new details (and colors).
Will test a bit with 2-pass if it can be made to be more smooth

Yep, second pass ruins it. Keep it as single. Also best to use a video editor when using this workflow.

Edit and cut the last 5 seconds of the recently created video and continue from that for a new generated video. Merge/Join a 5 second talking part from the original video and add it at the front of that newly cut 5 second video beforehand to keep the voice. Once that's done, edit the new longer video and remove the 5 second talking part at the beginning and merge the rest with the first edited video.

Can all be done easily through a freeware program like avidemux. It doesn't take long. Seems to be the only way as far as I'm aware to make really long videos without loosing character or voice.

Stick with the single extend workflow. Don't use the multi extend. Unfortunately that one works horribly and will mess up the characters and audio.

Also use the new OmniNFT lora and the Licon-VBVR-I2V-Video Reasoning390K-R32 lora and have them set at 1 strength. These are needed in order to actually follow the new prompt. I found not using them in this workflow makes the character just stand there like an idiot and not do anything.

Also, make certain the reference setting in this workflow at the bottom is not set higher then the added video or it will not transition properly. Setting it higher then the video length will cause it to skip/cut scene. To keep a smooth transition between the old and new, set it 1 second below your added video.

Something else to keep in mind, the lower your reference number is, the faster the generation will be. Useful if the talking takes place near the end of the video. You don't need to make the reference as long if that is the case.

Works great in single-pass and thats really the mode for such a wf.
Its nice for something quick and easy ...

For more serious video editing, using external editor, and doing one by one extension is better.
That way you can do each extended part over and over until you get the one you like... and keep going from there

Thanks for all the comments, I've been busy with other things so I havent been able to refocus on this, but I should have some time this upcoming week. :)

Yep, second pass ruins it. Keep it as single. Also best to use a video editor when using this workflow.

Edit and cut the last 5 seconds of the recently created video and continue from that for a new generated video. Merge/Join a 5 second talking part from the original video and add it at the front of that newly cut 5 second video beforehand to keep the voice. Once that's done, edit the new longer video and remove the 5 second talking part at the beginning and merge the rest with the first edited video.

Good advice for sure when it comes to keeping voice and motion consistency, but how are you dealing with frame jumps on the final merge? Extending the same video twice with this method and then splicing them together won't be seamless, as the last frame never aligns perfectly with the next starting frame.