Try using the LTXV audio text encoder node if you are having issues with Gemma!! #106
Description
If Gemma-3 causes LTX workflows to run extremely slowly, barely use the GPU, throw OOM errors, or appear to hang during prompt processing, the fix is to stop using the official Gemma / HuggingFace loader entirely and switch to the ComfyUI template method. Do not clone Gemma repositories from HuggingFace or rely on Git-LFS folders, as this often leaves pointer files and missing tokenizer assets that silently break performance. Instead, place a single-file Gemma model such as gemma_3_12B_it_fp4_mixed.safetensors directly into ComfyUI/models/text_encoders/ and load it using the LTXAVTextEncoderLoader (or LTXV Audio Text Encoder Loader) node. Then connect that node’s CLIP output to all CLIPTextEncode nodes in the workflow. This approach bypasses tokenizer/config issues, avoids fallback execution paths, dramatically improves speed, and restores normal GPU utilization. The fix applies equally to Text-to-Video, Image-to-Video, and F2F workflows. The LTX model itself is not the problem—the Gemma loader and packaging are. Looks like LTX didn't package their stuff properly to be used in comfyui without breaking.