LTX2.3 CLIP (CRT AutoDL)

LTX 2.3 is the speed tier of local video generation - the family you reach for when you want a draft clip in seconds instead of minutes - and its text encoder is a 12B Gemma 3 model. That's a lot of weight just to understand your prompt, and it's exactly the file people balk at downloading by hand. CRTAutoDLLTX23CLIP is the AutoDL loader that fetches both encoder files and gives you a ready CLIP output with no setup.

What makes this one interesting is that it isn't a single file. LTX 2.3 pairs Gemma 3 (the gemma_3_12B_it_fp4_mixed repack from Comfy-Org's LTX-2 repo) with a separate text projection head (ltx-2.3_text_projection_bf16 from Kijai's LTX2.3_comfy). CRT's loader knows this, downloads both into ComfyUI/models/text_encoders/, and loads them together with ComfyUI's LTXV CLIP type. You never have to remember which projection goes with which Gemma - the node does.

Using it

Like every AutoDL loader, it has no inputs. You wire the CLIP output into an LTX 2.3 workflow - the CRT pack's own LTX 2.3 Unified Sampler accepts it directly, and so does the standard conditioning path if you're building from core ComfyUI nodes. Pair it with CRTAutoDLLTX23Model and the LTX23 VAE nodes and you have the whole stack from six nodes.

The first run downloads ~12GB between the two files (Gemma at fp4 is surprisingly svelte for a 12B, but it's still a big pull). There's a console progress bar via CRT's download utility; the canvas shows nothing, so don't touch anything until you see CRT AutoDL activity in the terminal.

Gotchas worth knowing

• Disk, not VRAM. The fp4 Gemma plus fp8 model plus VAEs is a chunky install. LTX 2.3 is genuinely heavy on disk even though it's light on VRAM at inference.
• Don't substitute a full-precision Gemma. The loader expects the exact filenames it ships with. If you've already got a different Gemma build in text_encoders, it won't be picked up - and that's by design, because a mismatched projection head gives you garbage conditioning.
• Check you're actually on 2.3. This is the LTX 2.3 encoder; it will not load LTX-2 or LTXV checkpoints correctly.

Install the pack once via ComfyUI Manager (CRT-Nodes) or clone https://github.com/plugcrypt/CRT-Nodes.git into custom_nodes, run pip install -r requirements.txt, restart, and the node takes care of the rest on first queue.