LTX2.3 Model NVFP4 (CRT AutoDL)

LTX 2.3 is a 22B model. The fp8 build is a chunky ~22GB and the official recommendation was always beefy cards. LTX2.3 Model NVFP4 (CRT AutoDL) is the pack's answer for the rest of us: it downloads the NVFP4-quantized transformer - roughly half the size of fp8 - into models/diffusion_models, loads it, and hands you a MODEL ready for the LTX 2.3 unified sampler.

NVFP4 is NVIDIA's native 4-bit floating-point format, and it's the reason this node exists: RTX 40/50-series cards with fp4 support can run the 22B distilled model on far less VRAM than the fp8 path demands. Community reports have people running LTX 2.3-class models this way on 12–16GB cards where the fp8 weights wouldn't fit comfortably. If you're on an older card without good fp4 support, the GGUF Q5 variant in this same pack is the fallback worth reading next.

What the node actually does

It's an "AutoDL" loader: the CRT pack checks a registry of model definitions, and if ltx-2.3-22b-distilled_transformer_only_NVFP4.safetensors (from Winnougan's LTX-2.3-INT8 repo) isn't in diffusion_models, it streams it down with a progress bar on first use, then loads it through ComfyUI's normal diffusion-model path. The download happens on first execution, not at install time.

The three inputs are loader-level controls you mostly leave alone:

• patch_cublaslinear - toggles Comfy's cublas_ops fast-path flag. Leave off unless you're chasing specific matmul behavior on a card that benefits.
• sage_attention - here's the one worth understanding. The dropdown lists eight attention modes, but in the shipped code only two actually do something: disabled (stock attention) and auto (swaps in sageattention's kernel if it's installed). The specific kernel names in the list will raise an error if selected. So: leave it on auto if you've installed sageattention, disabled otherwise.
• enable_fp16_accumulation - default on; toggles allow_fp16_accumulation for fp16 matmul accumulation. Fine to leave on.

Output is a single MODEL, which you wire into LTX 2.3 US Models Pipe (CRT) alongside the video VAE, audio VAE, CLIP, and whatever mode-specific LoRAs your workflow needs.

Install and gotchas

It's part of the pack:

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

or ComfyUI Manager → search "CRT-Nodes". Gotchas, in order of how often they bite:

• The first run triggers the download, and it's a multi-gigabyte transfer. Don't panic when the node looks stuck - the progress bar prints to the console.
• sage_attention is the trap. The eight-mode dropdown oversells the shipped behavior. auto requires the sageattention package installed separately (pip install sageattention); without it, auto will fail on import, so use disabled unless you've set that up.
• NVFP4 quality is genuinely good for video - 4-bit on a 22B DiT is a much smaller hit than the GGUF quality ladder would suggest, because the format is designed for fp4 hardware. If it fits your card, it's the quant I'd reach for before the GGUF variants.
• Remember this loads only the transformer. You still need the video VAE, audio VAE, and text encoder - the pack's other AutoDL nodes or LTX 2.3 AutoDownload (CRT) panel cover those.