LTX-2-comfy_gemma_fp8_e4m3fn

An fp8_e4m3fn conversion of the text encoder at https://huggingface.co/Comfy-Org/ltx-2/blob/main/split_files/text_encoders/gemma_3_12B_it.safetensors, which is used for LTX-2

gemma_3_12B_it_fp8_e4m3fn.safetensors - The fp8 converted text encoder from comfy, goes in CLIP folder
gemma_3_12B_it_nvfp4_uncalibrated.safetensors - The nvfp4 converted (using comfyui kitchen utils) text encoder from comfy, goes in CLIP folder. NOTE: You will get warnings in the console, these can be ignored and will not affect your generation.
ltx-2-19b-dev-fp4_projections_only.safetensors - Extracted projections from LTX-2 model to allow loading with DualClipLoader node, goes in CLIP folder
ltx-2-19b-dev-fp4_video_vae.safetensors - The video vae, can be loaded with VaeLoader node, goes in VAE folder
ltx-2-19b-dev-fp4_vocoder.safetensors - The vocoder model, not useful separately currently
For audio vae, use Kijai's audio vae upload LTX2_audio_vae_bf16.safetensors, the one in this repo only contains the vocoder, but the audio vae is also needed. There was an oversight due to ComfyUI caching when it shouldn't.

Usage

When using a ComfyUI workflow which uses the original fp16 gemma 3 12b it model, simply select the text encoder from here instead.

Right now ComfyUI memory offloading seems to have issues with the text encoder loaded by the LTX-2 text encoder loader node, for now as a workaround (If you're getting an OOM error) you can launch ComfyUI with the --novram flag. This will slightly slow down generations so I recommend reverting this when a fix has been released.
--lowvram isn't needed if using the DualClipLoader, as it can be set to use cpu only.

Usage with DualClipLoader (with projections)

Use the vanilla ComfyUI DualClipLoader node, and as the clip models select the gemma_3_12B_it_fp8_e4m3fn.safetensors and ltx-2-19b-dev-fp4_projections_only.safetensors from this repo. Now replace the LTXV Audio Text Encoder Loader node with the DualClipLoader node.