The LTX-2-dev model yielded very poor results.

[juntaosun]

I tried using LTX-2-19b-dev + (distilled Lora), but the generated video is poor; it seems Lora isn't working?

I opened the workflows for three MP4 videos in the repository, and they also have Lora loaders commented out.

Does this mean the current node can only use the ltx-2-19b-distilled model and not LTX-2-19b-dev + (distilled Lora)?...

When will the LTX-2-19b-dev+ (distilled Lora) method be supported?

[RandomInternetPreson]

Greetings! Thank you for the screenshots. I think I understand what might be happening now regarding LoRAs and the node. I've tested the node with safetensors not ggus. I got reports that ggus worked, but now I'm wondering if ggus might work but not with LoRAs at the moment.

Can you do a few things for me (all fixed seeds)?

1. Just double check that with my node turned off you get the results you expect (to rule out an issue with node).
2. If you get good results with the node turned off, can you provide the most stripped down version of the workflow that results in a poor output? I'm not very good with ComfyUI and I need the most basic version of a workflow without any extra nodes.
3. With the workflow can you provide URL links to the specific model and LoRA

If you can provide this, I will set up the exact same workflow on my machine and do some debugging to see what the issue is.

[JonkoXL]

Hey RandomInternetPreson!

I also looked into this issue and it seems that any lora is just not being applied / or ignored when using this Node:
Tensor Parallel V3 (Safe FFN Chunking)

I attached two workflows, based on your Distilled V3 workflow, where I added two "LoraLoaderModelOnly" nodes to apply a LTX-2 pixel art style lora. (yes, apparently I used two lora nodes in sequence but that seems to apply the lora effect for me)
One workflow has your Tensor Parallel V3 node enabled and the other disabled.
Respectively I added two videos, one made with and the other without the Tensor Parallel V3 node enabled.

Distilled Disabled:
LTX-2_00003_Disabled.json
Video:
https://github.com/user-attachments/assets/c17a336d-0268-4f70-b461-58bb07570800

Distilled Enabled:
LTX-2_00003_Enabled.json
Video:
https://github.com/user-attachments/assets/6eea3978-c3b6-45b0-bcba-eefddc8e3fe8

The lora can be found here:
https://huggingface.co/datasets/JokerSmoker/Loras/blob/main/ltx2_pixel_v2_p1x3l_000001600.safetensors

If I had to make an educated guess, I think it's that any effect that the LORA applies to the model unfortunately gets lost somewhere, or is never applied.
I can be wrong but with WAN I remember loras sometimes being patched to models and then that patched model (maybe converted in memory or something) and then get used further along in the process. I also believe WAN VACE module patching uses something like this.

I also tried adding the Tensor Parallel V3 node to the default ComfyUI LTX-2 T2V template workflow, which uses the DEV model and not the DISTILLED model, but it distorts the video likewise.
I checked, and it's due to that template using the DEV model yet requiring a "distilled" lora as the poster above pointed out.
When I disabled adding that distilled lora, I get the same broken video.

ComfyUI Template Dev Disabled:
LTX2_T2V_ComfyTemplate_Disabled.json
Video:
https://github.com/user-attachments/assets/93ebb68b-e323-4362-a221-f719e230f7b1

ComfyUI Template DEV Enabled:
LTX2_T2V_ComfyTemplate_Enabled.json
Video:
https://github.com/user-attachments/assets/7c4b4d6e-71da-463a-bd99-da5205a85a44

ComfyUI Template DEV Disabled + Distill Lora disabled:
LTX2_T2V_ComfyTemplate_Disabled_AndNoDistillLora.json
Video:
https://github.com/user-attachments/assets/b77f8ec2-4df5-4057-a7f4-f2b6511d82b0

Note that everywhere I use the FP8 Safetensors models, so the issue should not necessarily be related to GGUF models.
The solution would probably be applying /patching the "lora" to the model that you use in your optimization code.
Maybe for getting it to work it's easiest if you make a field to load lora's in your own node, but I'm no expert :)

Either way hopefully the workflow examples + the lora can help you with debugging or looking into loras being applied correctly.

And thank you for being awesome, I am extremely happy with the performance improvements you made!