ltx-2.3_text_projection_bf16.safetensors

ltx-2.3_text_projection_bf16.safetensors is the projection layer that bridges the Gemma 3 12B text encoder and the LTX 2.3 DiT transformer. Gemma produces text embeddings in its own hidden dimension; the transformer's cross-attention expects a different dimension. This file is the learned linear projection that converts one to the other. Without it, the text path is dimensionally mismatched and the prompt never reaches the model.

It is a tiny BF16 component (~0.5 GB) and goes in ComfyUI/models/text_encoders/ — the same folder as the Gemma file (gemma_3_12B_it_fp4_mixed / fp8_scaled / comfy_gemma_3_12B_it). Kijai's ComfyUI port splits the text encoder into the Gemma weights plus this projection so each can be quantized and swapped independently.

This is not the text encoder itself and not a VAE. It carries no semantic knowledge — it is pure plumbing between two components. You still need a Gemma 3 12B file for the actual encoding.

Download this whenever your workflow loads the Gemma text encoder as separate components (the common Kijai ComfyUI setup). In that layout it is required — pair it with one Gemma 3 12B file: FP4 mixed on 16 GB, FP8 scaled on 24 GB, or the full BF16 on 32 GB+.

You can skip it only if your workflow uses a bundled/all-in-one text-encoder node that already contains the projection. If a text-encoder node turns red complaining about a missing projection or a shape mismatch, this is the file you are missing.

There is no quality choice to make here — it is BF16 only and the same file regardless of which Gemma quant you pair it with.

Some published workflow JSONs reference this file under a custom subdirectory. If ComfyUI shows a "cannot find model" error and your workflow references one of these path-prefixed variants:

• text_encoders/ltx-2.3_text_projection_bf16.safetensors
• ltx23\ltx-2.3_text_projection_bf16.safetensors

The prefix before the slash or backslash is a subdirectory the workflow author used. The actual file is the same ltx-2.3_text_projection_bf16.safetensors — you have two fixes:

1. Create the matching subdirectory inside ComfyUI/models/text_encoders/ and place the file there. Example: if the workflow references text_encoders/ltx-2.3_text_projection_bf16.safetensors, create the corresponding subfolder under ComfyUI/models/text_encoders/ and put ltx-2.3_text_projection_bf16.safetensors inside it.
2. Or open the workflow JSON in a text editor and replace the prefixed string with just ltx-2.3_text_projection_bf16.safetensors. ComfyUI then resolves it directly from ComfyUI/models/text_encoders/.

On Windows the separator is \, on macOS/Linux it is / — they refer to the same nested folder regardless of platform.