LTX2.3 Model GGUF Q5_K_M (CRT AutoDL)

LTX 2.3's 22B model doesn't fit on a 12GB card at fp8 - full stop. LTX2.3 Model GGUF Q5_K_M (CRT AutoDL) is the node that makes it fit anyway: it downloads unsloth's Q5_K_M GGUF of the distilled 1.1 model into models/unet_gguf, loads it through ComfyUI-GGUF, and hands you a working MODEL. The whole LTX 2.3-on-12GB community workflow scene - "LTX 2.3 GGUF 12GB workflows" is a whole genre of posts - runs on exactly this kind of quant.

Q5_K_M sits at the sweet spot of the GGUF quality ladder: the KB's GGUF panel puts Q5 as "the last stop before visible loss," with Q8 matching fp16, Q6 nearly identical, and Q4_K_M the accepted compromise. Q5_K_M is the point where most people can't tell the difference from fp8 but the weights actually fit. If you have a 12–16GB card, this is the node to reach for.

How it works

It's an AutoDL loader with a difference. The base AutoDL loaders call ComfyUI's native model loading; this one shells out to ComfyUI-GGUF (city96's pack) to do the actual dequantize-and-load. Specifically:

• it locates the ComfyUI-GGUF module in your custom nodes (and if it's missing, it raises a clear ImportError telling you to install it - that's not optional, it's a hard dependency),
• downloads ltx-2.3-22b-distilled-1.1-UD-Q5_K_M.gguf from unsloth/LTX-2.3-GGUF into unet_gguf on first run,
• loads it with GGMLOps as the custom operations set and wraps it in GGUF's model patcher.

Beyond the standard patch_cublaslinear, sage_attention, and enable_fp16_accumulation controls, you get the GGUF-specific knobs: dequant_dtype and patch_dtype (which dtypes the dequantized and patched layers use - default, target, or explicit float32/16/bfloat16) and patch_on_device (whether LoRA-style patching happens on-device). For plain sampling you can leave all four alone; they exist for when you're stacking LoRAs or chasing specific speed/VRAM behavior.

The sage_attention caveat

Same trap as every CRT AutoDL loader: the sage_attention dropdown lists eight modes, but the shipped code only acts on disabled and auto. auto imports sageattention - if the package isn't installed, it fails. The named kernel modes in the list raise an error if selected. So the practical choice is: disabled, or auto if you've actually installed sageattention.

Install

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

and separately:

cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF

Then restart. ComfyUI Manager search "CRT-Nodes" covers the first half; GGUF you may need to install manually.

Gotchas

First run downloads the GGUF (multi-GB), so expect a wait with console progress. GGUF pays a dequantization overhead - Q5_K_M is a touch slower than the NVFP4 path on fp4-capable cards, which is the trade you're accepting for VRAM. And because this loads only the transformer, you still need the LTX 2.3 video VAE, audio VAE, and text encoder from the pack's other nodes or the AutoDownload panel. Q5_K_M is genuinely good, but if your card supports fp4, the NVFP4 variant in this pack is faster and roughly as compact.