LTX2.3 Model (CRT AutoDL)

LTX 2.3 is a 22B distilled video model that still runs on consumer VRAM because the version you want is fp8-scaled. Getting that file from the right HuggingFace repo is the fiddly part - this node is the shortcut. CRTAutoDLLTX23Model downloads the distilled 1.1 fp8 transformer from Kijai's LTX2.3_comfy repack on first use, loads it as a ComfyUI MODEL, and - unlike a plain loader - gives you a few speed knobs while it's at it.

The three inputs that matter

Most AutoDL loaders are no-input convenience nodes. The model loaders are the exception, and here's what the three knobs do:

• sage_attention - when set to auto (the default), it swaps in SageAttention for the attention math if the sageattention package is importable, which is the biggest free speedup for big transformers on CUDA. If you don't have it installed, the loader still works - it just won't take the fast path. Other options range through the int8/fp8 SageAttention variants and the newer sageattn3 modes.
• enable_fp16_accumulation (default on) - toggles fp16 accumulation in cuBLAS matmuls. Leave it on unless you're chasing edge-case precision issues; off is measurably slower for most workloads.
• patch_cublaslinear - off by default. When on, it enables ComfyUI's --fast cublas_ops patch on linear layers, a real but modest speedup on some cards.

For a first run: leave all three at default, confirm it samples, then experiment with SageAttention variants once you've seen the baseline.

Wiring it

MODEL out goes to your sampler - CRT's LTX 2.3 Unified Sampler is the intended partner, and it'll take the standard core KSampler path too. You'll also want CRTAutoDLLTX23CLIP and the LTX23 VAE nodes to complete the stack. Output is a standard MODEL object, so nothing downstream knows or cares how it was loaded.

What to watch for

• The first queue downloads ~24GB. That's the fp8 22B transformer; there's a progress bar in the console and nothing on the canvas, so be patient. Plan around disk - this is the single biggest file in the whole CRT pack.
• sage_attention = auto is not a guarantee. It's "use SageAttention if available." If you pick a specific mode like sageattn_qk_int8_pv_fp16_cuda and it's not installed, the node raises rather than silently ignoring you - install sageattention (and, for triton modes, the matching triton build) or drop back to auto/disabled.
• This is the distilled model. It's the fast, 1-step-friendly variant of LTX 2.3 - the right default for most people, but know that it's not the full-fidelity base.

Install CRT-Nodes through ComfyUI Manager (search CRT-Nodes) or git clone https://github.com/plugcrypt/CRT-Nodes.git into custom_nodes, then pip install -r requirements.txt and restart. The model download kicks off on first queue.