Example workflow causing oom, not the model !

[john2stai]

At first, I thought it was the 0.9.1 model that cause the oom, but I tried 0.9.0 in the example workflow and still got oom. I then used 0.9.1 with the old comfy native workflow with stg from LTXTricks, it worked without any error.

[YAY-3M-TA3]

The reason is because the OOM is from the new VAE in the new node. The reason you do not see it in the native work flow is because the new VAE is NOT using the new VAE.

[john2stai]

The reason is because the OOM is from the new VAE in the new node. The reason you do not see it in the native work flow is because the new VAE is NOT using the new VAE.

I thought the latest comfy has support for the new VAE, so we can now use comfy native checkpoint loader to model and vae.

[x4080]

@john2stai using img2vid workflow from this repo I got OOM on SamplerCustomAdvanced, on regular comfy works fine, but using ltxtricks stg workflow I got error :

The expanded size of the tensor (16) must match the existing size (14) at non-singleton dimension 2. Target sizes: [1, 128, 16, 24]. Tensor sizes: [128, 14, 24]

How do you get the stg in lighttricks work ? can you share your workflow ?

Thanks

[LockMan007]

EDIT:

I now realize I misread the original post.
The new workflow uses a LLM for Prompt Enhancing or to get info about the image.
If you delete all the Prompt Enhancing stuff except the python String Function, I think it will use less resources and be faster.
I automatically removed this and never attempted to use it, because I didn't want more downloads and slower generations.

For TEXT to Video,
Remove:

• FluxPromptEnhance:
• String Replace (4 of these)

For Image To Video,
Remove:

• DownloadAndLoadFlorence2Model,
• Florence2Run
• String Replace (4 of these)

You can test this without saving the workflow and you can always save it as an alternative name or redownload the original workflows, so no reason to not try it.

---

---

Original message:
I also get the OOM and as people say to use the VAE, well you can't as far as I can tell.

I can't find a single workflow on the internet from searches of people using the new VAE with LTX video 0.9.1.
I downloaded the LTX 0.9.1 VAE.
https://huggingface.co/Lightricks/LTX-Video/tree/main/vae (The same repo where you download the 0.9.1 model)

"The new VAE is only supported in our Comfy nodes (https://github.com/Lightricks/ComfyUI-LTXVideo . If you use Comfy core nodes you will need to switch. Comfy core support will come soon."

They say you must use the LTX nodes, not native ComfyUI nodes to load the VAE, but there is not LTX VAE node to take the place to load the VAE, so if you use the ComfyUI "Load VAE" you get an error.

If you try to use a "beta Tiled VAE" node, it also throws errors with LTX 0.9.1.
I have updated all of these, from github directly, NOT using ComfyUI manager UpdateAll.

• Lightricks/ComfyUI-LTXVideo
  (not to be confused with Lightricks/LTX-Video)
• ComfyUI
• ComfyUI Manager
• Video Helper Suite

I am aware that using ComfyUI manager update all, will break things like SANA by installing a city96 version of the nodes which will output a black or gray image, so that is why I updated all these manually from github directly, to avoid any potential problems with comfyui manager getting the wrong files needed.

For troubleshooting in general for video models:
I recommend setting steps on 1 (minimum), frame_number on 9 (minimum) and resolution on 320x320.
Then you can try increasing one of these and see if it still works.

I have 11GB VRAM and 64GB System RAM and have it set to fallback on System RAM if it runs out of VRAM, but 0.9.1 doesn't seem to work with that or doesn't tile or something.

• LTX 0.9.0 works no problems.
• LTX 0.9.1 does work if I turn the size down to 320x320 with 15 steps, 89 length.
  and works at 512x512 at 15 steps, (length) 9 (any more, it gets an OOM.)

[x4080]

@LockMan007 where do you put the new VAE ?

[LockMan007]

@LockMan007 where do you put the new VAE ?

I can't get the LTX 0.9.1 custom VAE model to load in anything.

[x4080]

@LockMan007 Yes, I just put it in VAE folder and it seems dont do anything