TextGenerateLTX2Prompt

[RuneGjerde]

I see ComfyUI added the TextGenerateLTX2Prompt and a few other similar nodes.
Not sure where to post this, but since you made them. perhaps its ok here ;-)

Works great, and a workflow life-saver since i often use LLM to enhance prompt or describe images etc.

But one thing, it didnt seem to work with GGUF (and/or City69's dual GGUF clip loader), maybe thats something feasible to add.
(although i guess the Gemma gguf support is not truly part of ComfyUI support perhaps, but maybe KJnodes could)

And if not, or a lot of work to do, no biggie, the safetensor Gemma's works great ;-)

[kijai]

I would say it's up to the GGUF loader to fix as it's not something natively supported anyway, bigger issue is also that there are many GGUF models out there that do not include the vision and/or the multimodal projector weights at all.

That said, the unsloth 12B Gemma3 worked when I tested it for text generation at least.

[RuneGjerde]

yes its so many gguf files floating around, i might have tried one that was not ideal.
And no worries, works perfectly fine with safetensor.

(will try some unsloth ones though)

[kijai]

I also have this Gemma3 4B that includes vision weights up here for testing, it's far faster for general use:

https://huggingface.co/Kijai/Gemma3_comfy/blob/main/Gemma3_4b_it_vision_bf16.safetensors

Of course it can't be used to encode the actual LTX2 prompt embeds, but it can be used for image understanding and text generation otherwise.

[RuneGjerde]

yes safetensor is all ok. I was just thinking of those with extreme little vram, the Q2_K one might help.
But i see that comfy has mixed as well as fp4 safetensor for ltx-2, so its not that much you'd save on file size

And not sure it would even matter using Q2 with offloading etc, that comfy probably does ;-)

The text generate nodes are super nice though, can use the already loaded clip model to enhance prompt etc. Makes life easy ;-)

I also have this Gemma3 4B that includes vision weights up here for testing, it's far faster for general use:

yes, already downloaded that one, good for many workflows ;-)

No way this came out days after my LTX-2 Easyprompt node lol