Kijai

/

LTXV2_comfy

like 396

Download the main models from here:
https://huggingface.co/Kijai/LTXV2_comfy

• LTX-2 distil GGUF Q4 (models/unet folder)
• ltx-2-19b-embeddings_connector_distill (models/text_encoder folder)
• LTX2_audio_vae (models/vae folder)
• LTX2_vidoe_vae (models/vae folder)

Needed nodes:
https://github.com/kijai/ComfyUI-KJNodes (for loading split models)
https://github.com/city96/ComfyUI-GGUF (for loading GGUF models)

Download Gemma gguf model from here: https://huggingface.co/unsloth/gemma-3-12b-it-GGUF/tree/main (for example Q2 model) (models/text_encoder folder)

Then use these nodes :

• Unet GGUF model loader for LTX main model
• DualClip GGUF loader for Gemma GGUF and ltx-2 embeddings
• Vae loder for video vae
• KJVaeLoder for audio vae

That should hopefully work. Will post a workflow too, tweak it for low vram/ram

Try this workflow (8GB vram is very low, so no promises)
https://huggingface.co/RuneXX/LTX-2-Workflows/blob/main/LTX-2%20-%20T2V%20Basic%20(low%20vram).json

And setting the output resolution to a lower one might help too. For example 832x480
Here is one at 960x544 using the workflow above

(she didnt manage to say Vram, but thats on me, i should have prompted V-Ram or something hehe ;-))

Whats the minimum suitable V-RAM needed for this to run.

Whats the minimum suitable V-RAM needed for this to run.

not sure actually. Depends on comfy and offloading to ram, and how much ram you have etc
A wild guess is that 12-16GB might work (8gb really unsure, that might be too low)

Although some seem to have it running on 8GB
https://www.reddit.com/r/StableDiffusion/comments/1q87hdn/tips_on_running_ltx2_on_low_8gb_or_little_less_or/
(and with gemma gguf that he didnt have, maybe.. .;-) but the crucial part then is probably having enough regular ram and a huge windows page file (or automatic))

Someone was able to run LTX-2 with around 3GB VRAM usage by using --novram argument a few days after LTX-2 released.

Btw, are GGUF text encoders no longer need to be placed in models/clip folder anymore nowadays?

Download Gemma gguf model from here: https://huggingface.co/unsloth/gemma-3-12b-it-GGUF/tree/main (for example Q2 model) (models/text_encoder folder)

Btw, are GGUF text encoders no longer need to be placed in models/clip folder anymore nowadays?

text_encoders folder is also ok. I put it there since other workflows (and default comfyui workflow) put gemma in the text encoder folder

LTX is very vram friendly but it's hungry with RAM, at least 32GB RAM you need to make it run smoothly. I've been using LTX 2 with 6GB vram and 44GB RAM it's worked smoothly for 720p (1280*736) 10 sec.

here the presentation

https://youtu.be/Ph-jqonNvFo?si=m5I0RW9FCtG1VX-k

For WF i use native WF from comfyui but using KJ models and KJ nodes as well.

6GB of vram is impressive ;-) and yes, also my experience that this model (and/or optimisations done in comfy) runs beyond what i expected.

6GB of vram is impressive ;-) and yes, also my experience that this model (and/or optimisations done in comfy) runs beyond what i expected.

Yeah i also almost can't believe lol, such an advance model like LTX 2 can worked with my poor machine lol.