RuneXX

/

LTX-2-Workflows

like 235

Will make first frame → middle frame → last frame version of the K-sampler one.
It shouldnt complicate it too much ;-) since its really just 1 extra node + 3 image inputs

And yes, the whole workflow thing started due to a few posts i saw about people who couldnt easily figure out the default ones, that are "hidden" with nested sub-graphs etc.
So thought a more "plain" workflow might help some, while hiding some of the "spaghetti" logic and aiming for it to be easy to start with ;-)
As well as some might prefer it this way (me included...)

Hello, I hope you’re doing well.
May I ask for an additional request?

I’m currently using “LTX-2 - I2V and T2V Simple (1-pass K-Sampler).json” as my main workflow because it runs very reliably on my system. If possible, I would like to request adding audio recognition features to this workflow—specifically:

the MP3 voice / singing recognition functionality from “LTX-2 - I2V and T2V Basic (Custom Audio).json”, and

the MP3 voice recognition functionality from “LTX-2 - I2V Talking Avatar (voice clone Qwen-TTS).json”.

I tested those two workflows, but I often encounter OOM (out-of-memory) issues, and the output quality is sometimes not as good on my setup.
Because the Simple (1-pass K-Sampler) workflow works very well for me, I would really appreciate it if the audio recognition capability could be integrated into that workflow.

Thank you very much for your time and help.

Will do both ;-) the k-sampler workflow is probably one that many can prefer. Simple, easy and familiar.
(will probably add the sampler-preview too it as well , so you can see progress as the sampler runs.. Kijai made a nice node that made that possible)

Sorry, had a few hectic days with work.
The first - last - middle frame workflow for K-Sampler is uploaded now ;-)
At the LTXVAddGuideMulti node you can adjust the strength for each frame input.
(you can replace the guide node with LTXVImgToVideoInplace for a more strict frame, but personally i find it to be smoother with the guider node to give the model a little freedom)

One note is that its working a little less ideal on the last frame, but with a few runs you can get good results. Not sure if this is due to being just one sampler run. But will see if i can figure that one out.
(its like its rendering past last frame, for some reason. Probably since its a guider, not frame injector. You can set the position of last frame in the LTXVAddGuideMulti node. Setting it -12 etc seemed to work better)

I also added Kijai's sampler preview node so you can see the progress at the sampler (support for this might not yet be in the desktop version, but it will not prevent the workflow from running though)

Thank you so much for your help. I really appreciate it.

I’m working from home today, so I’ve only run the workflow a couple of times so far, but everything seems to be working well. I’m glad I reached out—you clearly know this area inside and out. 🙂

I do have one quick question: if I want to skip the mid-frame node, can I simply leave the mid-frame image empty and proceed? In other words, is it okay to provide only the first and last images?

And please don’t feel rushed—take your time. Thanks again!

I do have one quick question: if I want to skip the mid-frame node, can I simply leave the mid-frame image empty and proceed? In other words, is it okay to provide only the first and last images?

ComfyUI might throw and error if you just leave middle empty.

The easiest way would be at the LTXVAddGuideMulti to set the strength of the middle frame to 0. I havent tested, but in theory that should do it ;-)
Alternatively at the same LTXVAddGuideMulti you can change the inputs to 2, and reconnect first and last frame to the 2 inputs. The first frame and last frame value node is "hidden" beneath the LTXVAddGuideMulti node, so just drag it to the side to see the nodes to use ;-)

But I think setting the middle frame strength to zero will work fine ;-)

@RuneXX

its like its rendering past last frame, for some reason.

The guider appends the guide frames at the end of the generation. You need to use the LTXVCropGuides node on your video latent after the inference to trim them.

Good spotting ;-) will see if that works better.
Got a bit "distracted" with it just being one sampler, so i might have overlooked that one

UPDATE:
Checked and the workflow has the crop guides already.
Will see if its something else. ImageInPlace works more "correctly" per frame, but it flickers with that node. Will see if i can figure out either ;-)
But already works pretty well for just being one pass sampler, and K-Sampler, so maybe its ok as it is too ;-)

Checked and the workflow has the crop guides already.

That's really odd. because I had the exact same issue (I'm assuming) guide frames flashing at the end of the video, and adding cropGuides did the trick. I even went and checked the code for the guide node and it does say the guides are added as frames at the end.

https://github.com/kijai/ComfyUI-KJNodes/blob/main/nodes/ltxv_nodes.py#L46

Yes the K-Sampler workflow is just 1 sampler, and the "normal" thing is to add the crop in the 2nd pass upscale (that this workflow does not have).
It doesnt add more frames, its just a cosmetic thing, that it seems to run a bit past last frame (but it doenst actually have more frames than the set amount to render).

But will see if it anything ;-) It already works pretty well, it was just a minor thing.
Might also add the imageInPlace node for an alternative workflow - that is more exact frame injection (but it has a tendency to flicker a little bit)

I've found imageInPlace works better for the first frame. I think I'm doing something odd with the supported resolutions but on 9:16 aspect ratio I've had the guider do some weird stuff with my image, kinda squeezing it. I suspect some rounding errors in the way I resize my references where the guider thinks it's not actually divisible by 32.

If I had to guess, inplace for the first and last frame, with guider for the middle would probably give really good results.

If I had to guess, inplace for the first and last frame, with guider for the middle would probably give really good results.

Yes I have been thinking the same. But didnt add that in my workflows, since my aim was to keep things simple as possible (since a few got quite confused with the nested subgraphs in comfy and ltx default workflows).
But maybe the workflow spaghetti mess wont grow too much with both guider and in-place ;-) will give it a shot, as an alternative

RuneXX Is your model a girl wearing Iranian-style clothing, and are you yourself possibly from the great country of Iran?

RuneXX Is your model a girl wearing Iranian-style clothing, and are you yourself possibly from the great country of Iran?

Oh, I didnt know that ;-) i just thought it looked nice..
(the girl is Dua Lipa actually, but a photo of her that kinda didnt look like her, so I thought it would make a nice first frame. She is Kosovo Albanian descent, that is predominantly muslim, so might be connection there ;-)

I'm almost from the north pole.. hehe .. Norway ;-)