Kijai

/

LTXV2_comfy

like 214

same here, would appreciate if anyone can share a workflow

just the transformer, read the model card, there is an pending PR for City96 gguf loader, or just use the repo from the PR creator like I did it and copy the file over the comfyui-gguf nodes folder.

Same setup as the model card on this repro worked for me, but with GGUF loader instead of Diffusion model loader

BUT, its not yet supported though in the GGUF model loader. So you might want to wait a little, or pull the PR
https://github.com/city96/ComfyUI-GGUF/pull/399

EDIT: already answered above ;-)

(just a bit of a low res demo 720p, for fun ;-))

Hi @RuneXX , can you please share your workflow used for the above video.
Thanks

loadvaekj gave me this error when loading the audio. anyone know what caused this?
!!! Exception during processing !!! 'VAE' object has no attribute 'latent_frequency_bins'
Traceback (most recent call last):
File "D:\ComfyUI-Zluda\execution.py", line 518, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI-Zluda\execution.py", line 329, in get_output_data
return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI-Zluda\execution.py", line 303, in async_map_node_over_list
await process_inputs(input_dict, i)
File "D:\ComfyUI-Zluda\execution.py", line 291, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "D:\ComfyUI-Zluda\comfy_api\internal_init.py", line 149, in wrapped_func
return method(locked_class, **inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI-Zluda\comfy_api\latest_io.py", line 1582, in EXECUTE_NORMALIZED
to_return = cls.execute(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI-Zluda\comfy_extras\nodes_lt_audio.py", line 153, in execute
audio_freq = audio_vae.latent_frequency_bins
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'VAE' object has no attribute 'latent_frequency_bins'

Hi @RuneXX , can you please share your workflow used for the above video.
Thanks

You can download the video and drop into comfy. But keep in mind the GGUF model loader with LTX-2 metadata is not yet supported
https://github.com/city96/ComfyUI-GGUF/pull/399 So you need to pull the PR, or wait a little bit ;-)

@ryuunami1 Probably wrong VAE loader node, see here https://github.com/city96/ComfyUI-GGUF/pull/399
(but it might be you that also asked there ;-)

Use the loader from KJNodes, as shown on the model card on repro here: https://huggingface.co/Kijai/LTXV2_comfy
From : https://github.com/kijai/ComfyUI-KJNodes

@ryuunami1 Probably wrong VAE loader node, see here https://github.com/city96/ComfyUI-GGUF/pull/399
(but it might be you that also asked there ;-)

Use the loader from KJNodes, as shown on the model card on repro here: https://huggingface.co/Kijai/LTXV2_comfy
From : https://github.com/kijai/ComfyUI-KJNodes

HORRENDOUS result!! using GGUF Q4 LMFAO!!!

ANYONE knows were to find the COMFY TEMPLATES FOLDER so i can add the "nodes.py" and "loader.py" files to use GGUF?

HORRENDOUS result!! using GGUF Q4 LMFAO!!!

Worked fine for me. The video i posted above is made with GGUF.
Maybe something with your workflow

HORRENDOUS result!! using GGUF Q4 LMFAO!!!

Worked fine for me. The video i posted above is made with GGUF.
Maybe something with your workflow

im using YOUR workflow extracted from the video buddy.....

im using YOUR workflow extracted from the video buddy.....

Probably didnt pull the PR for gguf support?
its not supported yet, so you have to wait a bit ;-)

or pull the unreleased PR https://github.com/city96/ComfyUI-GGUF/pull/399

And if you dont know how to, just wait until its out in public version, it wont take long i bet ;-)

As its mentioned in front page of this repro

Give it a day or two and i bet support its out in the regular outdates ;-)
(before that it will not work without it)

im using YOUR workflow extracted from the video buddy.....

Probably didnt pull the PR for gguf support?
its not supported yet, so you have to wait a bit ;-)

or pull the unreleased PR https://github.com/city96/ComfyUI-GGUF/pull/399

And if you dont know how to, just wait until its out in public version, it wont take long i bet ;-)

As its mentioned in front page of this repro

Give it a day or two and i bet support its out in the regular outdates ;-)
(before that it will not work without it)

Yes, I got it working now with the TV news woman

But it won’t work with any other image

I got a very bad quality and the image won’t do lip sync

Bummer as I mentioned before I rather wait a whole month for this error to be fixed

But thanks for your help though

@ryuunami1 Probably wrong VAE loader node, see here https://github.com/city96/ComfyUI-GGUF/pull/399
(but it might be you that also asked there ;-)

Use the loader from KJNodes, as shown on the model card on repro here: https://huggingface.co/Kijai/LTXV2_comfy
From : https://github.com/kijai/ComfyUI-KJNodes

HORRENDOUS result!! using GGUF Q4 LMFAO!!!

I had result exactly like this, I used full model with low CFG and steps.
You need the distilled LORA for the upscaler if you use the normal model.

The noisy result happens when the model config isn't loaded from the metadata. Currently it means:

Applying this PR:
https://github.com/city96/ComfyUI-GGUF/pull/399
Or using their fork:
https://github.com/vantagewithai/Vantage-GGUF

Or using these nodes (I hear, didn't test myself):
https://github.com/calcuis/gguf

The noisy result happens when the model config isn't loaded from the metadata. Currently it means:

Applying this PR:
https://github.com/city96/ComfyUI-GGUF/pull/399
Or using their fork:
https://github.com/vantagewithai/Vantage-GGUF

Or using these nodes (I hear, didn't test myself):
https://github.com/calcuis/gguf

Thank you so much for your hard working !!!

and it's worked !!!! so awesome !!! even in my lower GPU (RTX 3050 6GB) it rendered very fast!! 485 sec for 5 sec 1280*720

here the results

here the ss of my GPU

cheers

got the distilled model to run but im encountering an issue with noise even though ive already applied the PR. videos workflow shows all my files im running so i dont know what the issue really is. could it be the lora running on top of the distilled model?

got the distilled model to run but im encountering an issue with noise even though ive already applied the PR. videos workflow shows all my files im running so i dont know what the issue really is. could it be the lora running on top of the distilled model?

you don't need distilled lora for distilled model, distilled lora needed just for dev/full versions.

can we use dualcliploader(gguf) with gguf version of gemma3 and lxt2 connectors to even reduce the memory requirements of this workflow? will it possibe? The only missing point seems to be there is no gguf connectors as for now.

can we use dualcliploader(gguf) with gguf version of gemma3 and lxt2 connectors to even reduce the memory requirements of this workflow? will it possibe? The only missing point seems to be there is no gguf connectors as for now.

I am trying to do the same thing with the Clip, use gguf for gemma, but it says, "Unexpected text model architecture type in GGUF file: 'gemma3'". I am using the pr mentioned, plus I also have the latest calcuis custom_node for gguf installed (but that seems like the functionality was added and then commented out, in version 2.79). Looking over the pr, it seems like the alteration targets the unet. I'm not sure if something similar needs to be done for the DualClipLoader or not.

hey where are the number of steps , im using gguf model

can we use dualcliploader(gguf) with gguf version of gemma3 and lxt2 connectors to even reduce the memory requirements of this workflow? will it possibe? The only missing point seems to be there is no gguf connectors as for now.

I am trying to do the same thing with the Clip, use gguf for gemma, but it says, "Unexpected text model architecture type in GGUF file: 'gemma3'". I am using the pr mentioned, plus I also have the latest calcuis custom_node for gguf installed (but that seems like the functionality was added and then commented out, in version 2.79). Looking over the pr, it seems like the alteration targets the unet. I'm not sure if something similar needs to be done for the DualClipLoader or not.

There are already working modifications to the GGUF custom node, I just got my working with GGUF Gemma3, was using this for help:
https://github.com/city96/ComfyUI-GGUF/pull/402

I needed to merge also #402 on top of #399 for the GGUF Gemma to work. Also for gemma I copied the tokenizer.model to the same subfolder I have my GGuFs at (not sure if strictly needed). But your torch and cu are quite old, not sure what the requirements are though.