RuneXX

/

LTX-2.3-Workflows

like 657

Hmm not sure. Is it perhaps a video input where no body is detected? for example some animated cartoon or something that the SD Pose doesnt "understrand"?

I've switched to SW and the video input is detected and saved as video pose so there is a problem with SD pose. But now with SW the rendered video is not performed. Just the pose video is saved. I'ive tried to bypass the second pass but no result. I don't see any switch. No error message. I have 32 GO RAM and 12 Go Vram with 512*768 definition set. Maybe this workflow is heavy. I'm going to try the gguf version. Error with gguf. 'RuntimeError: Error(s) in loading state_dict for LTXAVModel: size mismatch for audio_embeddings_connector.learnable_registers: copying a param with shape torch.Size([128, 2048]) from checkpoint, the shape in current model is torch.Size([128, 3840]).
Do you know which gguf version to use ?

Did you add the lora? The lora removes the pose from the video (and uses the pose to drive the video output)

(there is a preview video of the pose at bottom of workflow, but this is not save, just there to preview if the pose was detected)

realisdance vs IC-Union-Control who is better ?

I'm surprised there's not more buzz about the Realisdance workflow. I find the results to be rather impressive myself. Rune, would it be possible to please incorporate the Tenstrip facial anchor and tiled sampling improvements into the Realisdance workflow for better facial likeness retention and character integrity during motion (and maybe in other workflows where relevant)? AI Verse did a video on the Tenstrip nodes here (looks like a big improvement to faces!):

https://www.youtube.com/watch?v=Ikh5EZu8LNQ

The Tenstrip custom nodes and their provided workflows for facial anchoring with tiled sampling are here:

https://github.com/TenStrip/10S-Comfy-nodes

Thanks!

+1 i like the realisdance lora too ;-) the result seems a bit more natural than the control-union

Not sure what the "facial anchor is", but will take a look. It should be possible i bet

Awesome thanks!!! Yeah realisdance is totally being slept on. I have the exact same experience with it vs control-union. I'd argue it might be the best open-source control-vid motion transfer method including all wan options (Scail is the best one there I think, but realisdance is maybe even better). The "Ai Verse" vid I linked before gives a good high-level overview of the Facial Anchor (with video comparison proof of results) and Tiled Sampling benefits. If the vid is correct, these two features could potentially add significant improvements across many/most workflows.

Did you add the lora? The lora removes the pose from the video (and uses the pose to drive the video output)

(there is a preview video of the pose at bottom of workflow, but this is not save, just there to preview if the pose was detected)

for me, too, the process ends with extracting the pose. no errors, but nothing is sent to the sampler.

Did you add the lora? The lora removes the pose from the video (and uses the pose to drive the video output)

(there is a preview video of the pose at bottom of workflow, but this is not save, just there to preview if the pose was detected)

for me, too, the process ends with extracting the pose. no errors, but nothing is sent to the sampler.

is there a solution? I also create a pose, and the process ends there. I've reviewed all the nodes, and everything is fine, but after extracting the pose, the generation doesn't continue

Did you add the lora? The lora removes the pose from the video (and uses the pose to drive the video output)

(there is a preview video of the pose at bottom of workflow, but this is not save, just there to preview if the pose was detected)

for me, too, the process ends with extracting the pose. no errors, but nothing is sent to the sampler.

is there a solution? I also create a pose, and the process ends there. I've reviewed all the nodes, and everything is fine, but after extracting the pose, the generation doesn't continue

it turned out to be a custom audio file, and you have to insert any file there in order for the process to continue

it turned out to be a custom audio file, and you have to insert any file there in order for the process to continue

Yes, thats a drawback of ComfyUI and workflows that has multiple ways of working.
Even if you are not using something, say image in an I2V + T2V workflow you still have to load a dummy image.
And the custom audio still needs a file loaded, even if its not being used