LTX2.3 Model GGUF Q4_K_M (CRT AutoDL)

LTX 2.3 is a 22B model. Full stop, that's a lot of VRAM. The GGUF quantization scene existed for exactly this moment: quantize the weights so the thing fits. This node is the "give me the practical version" button - it auto-downloads Unsloth's ltx-2.3-22b-distilled-1.1-UD-Q4_K_M.gguf and loads it as your diffusion model.

Q4_K_M is the workhorse tier of the GGUF ladder: not the Q8 "basically fp16, why bother" tier, but the one that actually gets 12–16GB-class cards running a model that would otherwise need ~40GB of weights. The KB's GGUF panel calls it the accepted compromise, and for a 22B video model that compromise is the difference between "runs" and "doesn't run."

How it works

Under the hood it's a thin wrapper over city96's ComfyUI-GGUF - it locates that custom node pack, uses its GGUF ops and loader, and builds a MODEL from the quantized file. The first run downloads the ~10GB-ish GGUF from HuggingFace (unsloth/LTX-2.3-GGUF) into models/unet_gguf/ with a progress bar. Every subsequent run is instant.

What you actually tune are the GGUF-to-torch mapping knobs, which the pack exposes directly:

• patch_cublaslinear - toggles the cublas matmul ops for linear layers; a speed lever for some cards.
• sage_attention - dropdown with disabled, auto, and a set of explicit SageAttention modes (sageattn_qk_int8_pv_fp16_cuda, triton variants, sageattn3, …). auto uses SageAttention when available, falling back gracefully. SageAttention is the big speed win on modern NVIDIA cards; if you don't have sageattention installed, auto degrades rather than crashing.
• enable_fp16_accumulation - on by default; leave it unless you're chasing numerical edge cases.
• dequant_dtype / patch_dtype - how the quantized weights are dequantized and patched (default/target/float32/float16/bfloat16). Defaults are the sane choice; these are for the deep-end experiments.
• patch_on_device - whether patching happens on the GPU or CPU.

Output is MODEL, which feeds the same place any LTX 2.3 model does - the pack's Models Pipe / Unified Sampler.

Installing it

Two installs, not one. CRT-Nodes itself (Manager → CRT-Nodes, or clone + pip install -r requirements.txt), plus ComfyUI-GGUF from https://github.com/city96/ComfyUI-GGUF. The node literally raises an ImportError naming that URL if it's missing - that's your cue. Then restart, drop the node in, and let the download run.

Where people get burned

The most common failure is skipping ComfyUI-GGUF - the node can't exist without it. Second is expecting fp8 quality: Q4 is a real quantization, and with video models the softness shows in fine detail, exactly where LTX 2.3 already had a reputation for being soft. Use this tier to get running and iterate; if quality stalls and you have the VRAM, the pack's fp8 distilled model node is the upgrade path. And if SageAttention isn't installed, the explicit sage modes will error - use auto or disabled unless you know you have it.