RuneXX

/

LTX-2.3-Workflows

like 587

TypeError: TextGenerateLTX2Prompt.execute() got an unexpected keyword argument 'use_default_template'

Yes it seems like the latest comfyUI update broke the prompt enhancer. I am sure they will fix that asap ...
Should be possible to bypass it for now though. I had same error yesterday, due to the ComfyUI update, and disabled the prompt enhancer for now
Should be fixed in next update though.

https://github.com/Comfy-Org/ComfyUI/issues/13448

Also, is it normal that I often get a static image with chopped audio or an almost perfect video but then the last few frames have some cartoon FX?

No ;-) that sounds strange.
What model are you using? and make sure that you are using the upscale model v.1.1 that explicitly fixes the last frames glitches

Static image is a recurring challenge with LTX though. Try start the prompt with something like "Make this image come alive with fluid motion" and focus the prompt of sequence of actions to happen, and less describing the image (LTX can already see the image). Focus on things that should happen, and the camera.
Some images end up really really stubborn though, if so try a different seed also

Also, is it normal that I often get a static image with chopped audio or an almost perfect video but then the last few frames have some cartoon FX?

No ;-) that sounds strange.
What model are you using? and make sure that you are using the upscale model v.1.1 that explicitly fixes the last frames glitches

Static image is a recurring challenge with LTX though. Try start the prompt with something like "Make this image come alive with fluid motion" and focus the prompt of sequence of actions to happen, and less describing the image (LTX can already see the image). Focus on things that should happen, and the camera.
Some images end up really really stubborn though, if so try a different seed also

I guess it's LTX being LTX. Happens over a wide range of models and workflow. LTX can't really do FLF, it's more like FLF-ish where the last frame is guidance or a suggestion.

Static image is mostly on very short videos or a last frame that isn't different enough. LTX really wants a minimum length and a last frame that has enough difference.

Static image is mostly on very short videos or a last frame that isn't different enough. LTX really wants a minimum length and a last frame that has enough difference.

Yes this i can easily imagine being challenging, and the model seeing as a static or sliding image instead of a "video"

Btw, did you try FLF with the version 1.1 of the distilled model? (and v1.1 of the upscale model).
Only did 2-3 runs earlier, to test something else, but seem to remember the end frame was pretty accurate, at least in my tries, but might have been some luck
It is like you say, the model will take its freedoms (usually for the good, to make things more natural, but not always)

But if its something very specific like product "ad" or some very specific transition (unlike a human as in my video), it might be more noticable

Static image is mostly on very short videos or a last frame that isn't different enough. LTX really wants a minimum length and a last frame that has enough difference.

And try the transition lora. That might help force a bit of change (not tried your specific scenario, but worth a try)
Simply add the lora to the workflow, and add the trigger word at end of prompt ( trigger word: zhuanchang) . But i think it works without trigger word also
https://huggingface.co/valiantcat/LTX-2.3-Transition-LORA

Can see some runs i did here https://huggingface.co/RuneXX/LTX-2.3-Workflows/discussions/88

A transtion lora workflow here https://huggingface.co/RuneXX/LTX-2.3-Workflows/tree/main/First-Last-Frame
But if i recall correctly, its nothing special about the workflow, just the lora added

Btw, did you try FLF with the version 1.1 of the distilled model? (and v1.1 of the upscale model).

Yes. I specifically tried the workflow because 1.1 is supposed to get rid of the blur at the end of a video, which was annoying when you needed your video to continue.

FLF probably simply doesn't work well with a 1 second video. Didn't even work when I increase the frames to have more frames in the timeframe. Oh well.

Yes if you plan to continue the video, the last frame needs to be spot on. Will try see if there is any way to force that to happen.

hi,rune master,
I've encountered a problem: in the final stage of this workflow（LTX-2.3_-_FLF2V_First-Last-Frame workflow）, the audio output is showing an error. Here's the error message:

Video Combine 🎥🅥🅗🅢

Exception: An error occured in the ffmpeg subprocess:
[aac @ 0000026eb19554c0] Input contains (near) NaN/+-Inf
[aost#0:1/aac @ 0000026eb1f30a00] Error submitting audio frame to the encoder
[aost#0:1/aac @ 0000026eb1f30a00] Error encoding a frame: Invalid argument
[aost#0:1/aac @ 0000026eb1f30a00] Task finished with error code: -22 (Invalid argument)
[aost#0:1/aac @ 0000026eb1f30a00] Terminating thread with return code -22 (Invalid argument)；

Then, I skipped the audio compositing node, and the video appeared, but there was no sound. What could be the reason for this?

The ltx audio vae code in ComfyUI AND in KJNodes were both updated with new codes the other day.
Try update both ComfyUI and KJNodes and it should work again ;-)

ok ,thank you ；That should be something that happens today. I was fine yesterday. Thanks again.

Yes depends on when you updated comfy. The comfy code change broke the KJNodes vae loader.
But this was since fixed in a KJNodes update..

(and to not confuse, comfy was then also update another time, so now you can even load the audio vae in regular vae loader (and in KJNodes vae).. but both needs to be up to date ;-))