Why is this happening?
At first, i got a woman speaking,
but then all images stay static........
It seems to me that the model has an easier time with images sourced from actual video, like TV-Shows or movies.
I get those "stills" alot with photographs or AI images. But screenshots from The Sopranos clips, work well almost 100% of the time.
Just early impressions, could be wrong here.
It seems to me that the model has an easier time with images sourced from actual video, like TV-Shows or movies.
I get those "stills" alot with photographs or AI images. But screenshots from The Sopranos clips, work well almost 100% of the time.
Just early impressions, could be wrong here.
weird,,,,, you think has something to do with some DENOISE or STRENGHT in the nodes?
It seems to me that the model has an easier time with images sourced from actual video, like TV-Shows or movies.
I get those "stills" alot with photographs or AI images. But screenshots from The Sopranos clips, work well almost 100% of the time.
Just early impressions, could be wrong here.
but the second image i used is a REAL photo and the WORKING image is an AI image, lol
I remember the old LTX someone said adding some fake film grain/noise made things work better.
But if thats a myth or true, i dont know ;-) and if it apply to LTX-2 as well not sure
But i had a few static images too when doing I2V. Probably just try a different seed to fix that
It seems to me that the model has an easier time with images sourced from actual video, like TV-Shows or movies.
I get those "stills" alot with photographs or AI images. But screenshots from The Sopranos clips, work well almost 100% of the time.
Just early impressions, could be wrong here.
but the second image i used is a REAL photo and the WORKING image is an AI image, lol
Yes, but like I said, photos and AI images work poorly in my experience. Both of them.
However, this image for example, has never produced a still for me:
I have similar Sopranos screenshots and only one of them has ever produced stills. Success rate has been extremely high with them overall.
SOMETIMES it works fine:
sometimes it JUMPS to another person
what happens if you add this node between your input image and the rest..
Its been ages since i used the "old" ltx, but was some tricks there if i remember right
(And this node above is also present in LTX-2 before the upscale)
The theory being that if the image is too perfect (too hq), the model treats it like a poster. But thats just pure speculations ;-) i dont really know
Jump to a different person was a bit odd though...
what happens if you add this node between your input image and the rest..
Its been ages since i used the "old" ltx, but was some tricks there if i remember right
(And this node above is also present in LTX-2 before the upscale)
its connected already, not sure
its connected already, not sure
yes thats before the upscale part. I meant adding it also at the first input image.
Will try some here too ;-) if i see that static image happen again
Edit:
You are right. The compression of image is already there, for the input image as well. Hmm. Not sure then
do any of you guys have the TEXT to video version with kijais node already added?
do any of you guys have the TEXT to video version with kijais node already added?
Huggingface dont allow json attachment it seems, but drop the video into comfy
do any of you guys have the TEXT to video version with kijais node already added?
Huggingface dont allow json attachment it seems, but drop the video into comfy
ok thank you
Will try the LTX-Video workflow too, see if that works ;-) (its a bit different than the native)
KJ nodes on the LTX-Video workflow (a little different than the native workflow).
Tried a few runs, seems to work very well
I2V
The old LTX models I believe were mainly trained on images from actual videos, which included compression noise added by mpeg encoders a lot of the time - so the standing theory I think was that the model learned to associate that kind of compression artifacts and noise with motion.
Thusly was born the "tricks" where you added this type of compression noise/artifacts in different ways.
I assume that this is what that LTXVPreprocess node does or something similar - earlier this kind of effect was also achievable by just pre-processing the loaded image by saving it as a video with h264 codec using https://comfyai.run/documentation/VHS_VideoCombine and then extracting the image back and feeding it into the AI, this worked like 90% of the time to get it to properly animate a picture.
GUYS, where do i change the steps??? i cant find it..... i want to try from 8 steps to 12..... any help?
KJ nodes on the LTX-Video workflow (a little different than the native workflow).
Tried a few runs, seems to work very wellI2V
hello friend, how do i change the steps?
hello friend, how do i change the steps?
Think thats in some advanced sigma, "hardcoded"
But if you use the full non distilled workflow, and just change model to distilled and set CFG to 1, you can try using 8 to 12 steps etc. The full model workflow has a step node
Think that might be fine (but i didnt try that yet)
hello friend, how do i change the steps?
Think thats in some advanced sigma, "hardcoded"
But if you use the full non distilled workflow, and just change model to distilled and set CFG to 1, you can try using 8 to 12 steps etc. The full model workflow has a step node
Think that might be fine (but i didnt try that yet)
yeah i removed the sigma node and added the steps node!!! thanks!
LTXVEmptyLatentAudio
'VAE' object has no attribute 'latent_frequency_bins'
LTXVEmptyLatentAudio
'VAE' object has no attribute 'latent_frequency_bins'
Looks like you are trying to use GGUF model?
Its not yet fully supported, the GGUF model loader node needs an update.
Either wait a little bit with patience until its out in the public version, or use git to pull the pr
Although you have a VAE error, maybe you connected something wrong with the VAE loaders (if you have already the PR update from the GGUF model loader)
But thats beyond my "pay grade", not sure why, will leave that for the experts ;-)
hello friend, how do i change the steps?
Think thats in some advanced sigma, "hardcoded"
But if you use the full non distilled workflow, and just change model to distilled and set CFG to 1, you can try using 8 to 12 steps etc. The full model workflow has a step node
Think that might be fine (but i didnt try that yet)
hello mate!!! any idea why i get this result using YOUR video as WF? i ONLY change the model to NON distilled and nothing else was changed and i get these results1!!!
Not sure about that one, I also get "static" or slow zoom sometimes without the main subject talking. I havent quite figured out why that sometimes happens, or if its just the model, and try other seed
Not sure about that one, I also get "static" or slow zoom sometimes without the main subject talking. I havent quite figured out why that sometimes happens, or if its just the model, and try other seed
thanks..... ok have a great day,,, seems i have to wait some more WEEKS so all theses issues are resolved,,,, bummer....
Not sure about that one, I also get "static" or slow zoom sometimes without the main subject talking. I havent quite figured out why that sometimes happens, or if its just the model, and try other seed
thanks..... ok have a great day,,, seems i have to wait some more WEEKS so all theses issues are resolved,,,, bummer....
I agree... think we need to wait to see how this evolves because as is, it's not really useful to be honest. Was hoping for Wan 2.5-like capabilities, but AI-Santa did not deliver ;)
do any of you guys have the TEXT to video version with kijais node already added?
Huggingface dont allow json attachment it seems, but drop the video into comfy
I MADE A JOKE COMPILATION using your example video, LOL
Hello guys, I2V workflow anyone?
For "static image" try this:
https://huggingface.co/Kijai/LTXV2_comfy/discussions/1#6963cd90bd6f599dc4547fef
(at least it worked for me, not had that issue ever since)