LTX2 gemma 3 12B support

[zwukong]

LTX2 is supported by native comfyui

https://huggingface.co/Comfy-Org/ltx-2/blob/main/split_files/text_encoders/gemma_3_12B_it.safetensors
https://huggingface.co/unsloth/gemma-3-12b-it-GGUF/tree/main

[gitmylo]

And the ltx-2 repo links the QAT q4_0 unquantized version specifically so I feel like q4 would work perfectly, also ComfyUI uses a different text encoder loader for ltx2 which references the main model, not sure what it does exactly though, but something would probably need to be added to bridge that gap.

[MeiYi-dev]

Sadly it will take more time because of this Comfy-Org/ComfyUI#11684

[gitmylo]

Sadly it will take more time because of this Comfy-Org/ComfyUI#11684

I have a workaround, these work with the regular DualClipLoader nodes (loading gemma3 in one, the projections in the other). So gguf quants of these separated models should be compatible with the gguf DualClipLoader node.
https://huggingface.co/GitMylo/LTX-2-comfy_gemma_fp8_e4m3fn/tree/main

I'm also planning to separate the vae and vocoder and strip the separated parts from the main model to fully separate the models so they can be loaded with just the regular comfy nodes.

[MeiYi-dev]

Sadly it will take more time because of this Comfy-Org/ComfyUI#11684

I have a workaround, these work with the regular DualClipLoader nodes (loading gemma3 in one, the projections in the other). So gguf quants of these separated models should be compatible with the gguf DualClipLoader node. https://huggingface.co/GitMylo/LTX-2-comfy_gemma_fp8_e4m3fn/tree/main

I'm also planning to separate the vae and vocoder and strip the separated parts from the main model to fully separate the models so they can be loaded with just the regular comfy nodes.

I assume it would require some code changes/additions to ComfyUI for loading the VAEs, vocoder, etc as seperate files.

Their repo has seperated files now so no need for that https://huggingface.co/Lightricks/LTX-2/tree/main

[gitmylo]

I assume it would require some code changes/additions to ComfyUI for loading the VAEs, vocoder, etc as seperate files.

Nope, I've also separated the vae and the vocoder and just have to test those now. The vocoder still needs the audio vae loader node sure, but the vae should load just fine with the vae loader node

[MeiYi-dev]

I assume it would require some code changes/additions to ComfyUI for loading the VAEs, vocoder, etc as seperate files.

Nope, I've also separated the vae and the vocoder and just have to test those now. The vocoder still needs the audio vae loader node sure, but the vae should load just fine with the vae loader node

OOH! nice, hopefully it works. The "seperate nodes for loading only for this model" is triggering me so much XD

[MeiYi-dev]

I assume it would require some code changes/additions to ComfyUI for loading the VAEs, vocoder, etc as seperate files.

Nope, I've also separated the vae and the vocoder and just have to test those now. The vocoder still needs the audio vae loader node sure, but the vae should load just fine with the vae loader node

https://huggingface.co/GitMylo/LTX-2-comfy_gemma_fp8_e4m3fn/blob/main/ltx-2-19b-dev-fp4_projections_only.safetensors this should be fp8 since fp4 only works for blackwell GPUs no?

[gitmylo]

OOH! nice, hopefully it works. The "seperate nodes for loading only for this model" is triggering me so much XD

It works, I'll upload those files and then see if I can also save the transformer after

https://huggingface.co/GitMylo/LTX-2-comfy_gemma_fp8_e4m3fn/blob/main/ltx-2-19b-dev-fp4_projections_only.safetensors this should be fp8 since fp4 only works for blackwell GPUs no?

Comfy kitchen is a workaround for supporting nvfp4 on non-supported platforms with dequant. 50 series get a speedup but other gpus should be able to run it as well now.

[MeiYi-dev]

OOH! nice, hopefully it works. The "seperate nodes for loading only for this model" is triggering me so much XD

It works, I'll upload those files and then see if I can also save the transformer after

https://huggingface.co/GitMylo/LTX-2-comfy_gemma_fp8_e4m3fn/blob/main/ltx-2-19b-dev-fp4_projections_only.safetensors this should be fp8 since fp4 only works for blackwell GPUs no?

Comfy kitchen is a workaround for supporting nvfp4 on non-supported platforms with dequant. 50 series get a speedup but other gpus should be able to run it as well now.

Yeah but a big server I am inside people are saying fp4 is producing bad results for them. https://discord.gg/b6JphDpr here if you want to see :)

[gitmylo]

The parts uploaded here aren't really fp4 though mostly. Should work fine

[MeiYi-dev]

The parts uploaded here aren't really fp4 though mostly. Should work fine

Thanks for this!

[MeiYi-dev]

The parts uploaded here aren't really fp4 though mostly. Should work fine

Kijai is working on this (the audio VAE part). He just needs permission from comfy for it to be commited.

[gitmylo]

Yeah that would allow it to be loaded using the regular vae loader (although I'd need to save it in a different format probably).

Currently this is how you use the vocoder (audio vae) from the huggingface repo, pretty much identical except you have to use the LTXV Audio VAE Loader

[MeiYi-dev]

Yeah that would allow it to be loaded using the regular vae loader (although I'd need to save it in a different format probably).

Currently this is how you use the vocoder (audio vae) from the huggingface repo, pretty much identical except you have to use the LTXV Audio VAE Loader

Though Kijai is working on the audio vae part, not the vocoder.

[gitmylo]

The vocoder is the audio vae, it's just a different word used for the same thing