LTX2.5 CLIP w4a8 Light (CRT AutoDL)

The text encoder is the hidden VRAM sink in every modern video model, and LTX2.5's is a Gemma 4 12B. LTX2.5 CLIP (CRT AutoDL) loads a w4a8-quantized version of it - 4-bit weights, 8-bit activations - so the thing that used to eat 22+ GB in fp16 costs a fraction. No inputs, one CLIP output, and the checkpoint downloads itself on first run into ComfyUI/models/text_encoders/.

This is the "Light" half of the LTX2.5 memory story. The pattern is familiar from the LTX-2.3 era, where the Gemma 3 12B encoder caused most of the OOM errors at launch and quantized encoders became the standard fix. CRT's take is a w4a8 Light quant from the same LTX-2.5-Quantized repo that supplies the model quants, so it slots into the AutoDL stack cleanly.

Why the encoder matters

It's tempting to treat this node as a boring loader - and it is, mechanically - but it's the difference between an LTX2.5 workflow that fits and one that doesn't. The encoder loads first and sits resident while you sample, so its footprint is overhead on top of the DiT. A 12B encoder at w4a8 plus a 13–17 GB model quant is how the 16 GB tier stays viable. If you try to run LTX2.5 with a full-precision 12B encoder, you'll wonder why your 24 GB card is paging.

Wire the CLIP output into your CLIP Text Encode, alongside the LTX2.5 Model, VIDEO VAE, and AUDIO VAE from the same CRT stack. It's the w4a8 variant by design - there's no "full precision" option in this node, because for this model that option is mostly a mistake on consumer hardware.

Installing it

Standard CRT install:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt

or ComfyUI Manager → search CRT-Nodes, restart. First run downloads the encoder (still a few GB even quantized), with a console progress bar.

Common issues

• First run slow - the download. One-time.
• Wrong encoder for the model - LTX2.5 needs its own Gemma-based encoder; the CLIP from another LTX generation won't encode for the 2.5 DiT correctly. Stick to the CRT LTX2.5 set.
• VRAM still tight - the w4a8 encoder is already the memory play; next lever is attention_method on the model node (SageAttention) or dropping a model tier.
• Update breakage - red socket or NaN → right-click → Fix node (recreate).