LTX2_SM_Model

Every LTX-2.3 workflow in this pack starts here, and honestly the name undersells it. This isn't just "load a model" - it's the node that decides how the model runs: which checkpoint, which quant, whether the distilled LoRA is stacked on, and which of the many LTX-2.3 sampling modes the pipeline assembles. It's also the reason you can run a 22B audio+video model on a 6GB card at all, which is the whole pitch of this pack.

What it is

LTX-2.3 is Lightricks' 22B DiT - the one open model that generates video and synchronized audio in a single pass, and the model the community largely anointed the best thing you can run locally by mid-2026. The official ComfyUI native path exists and is fine on big cards. This pack is the "I have a small GPU and a lot of RAM" lane: smthemex (a wrapper author who does this for every new model release, GGUF-first) vendored Lightricks' own pipeline code and wired it to load quantized weights with streaming offload. The author's README calls it a test node. Treat it as that: a working, community-maintained wrapper that moves fast.

Under the hood, load_model() picks either your dit (a .safetensors from models/diffusion_models) or your gguf (a quant from models/gguf) - you choose one, the other stays "none". If the filename contains "fp8" it auto-applies fp8 quantization. Then it stacks the distilled_lora at a fixed strength of 0.6 using Lightricks' Comfy-renaming map, plus your lora at 1.0. Set offload on (default) and the model lives on the CPU and streams layers to the GPU as needed - that's the 6GB trick.

Inputs that matter

• dit / gguf - the transformer, one or the other. The README's files are ltx23-transformer-distill-Q8_0.gguf (or the 1.1 revision / Q6_K) for GGUF, or ltx-2.3-22b-distilled.safetensors as a non-GGUF dit.
• distilled_lora - the LTX-2.3 distilled LoRA. Use 1.1, not 1.0; 1.0 mumbles in the first sampler stage. It's ~7GB and sits in models/lora.
• sampling_mode - the dropdown that changes the whole pipeline: distilled, one_stage, two_stages, twostages_hq, keyframe, ic_lora, audio2v, retake. Pick distilled for 8-step CFG-1 generation; ic_lora when you're running an InContext LoRA; audio2v for audio-driven video.
• offload - leave on unless you have 24GB+ VRAM and want speed.

The single output is model, which feeds the KSampler.

Installing it

The pack installs like any custom node - via ComfyUI Manager (search "ComfyUI_LTX2_SM") or:

cd ComfyUI/custom_nodes
git clone https://github.com/smthemex/ComfyUI_LTX2_SM.git
cd ComfyUI_LTX2_SM
pip install -r requirements.txt

Restart ComfyUI. The requirements are the real story: diffusers, accelerate, transformers, tokenizers, gguf, omegaconf. Note this loads through HuggingFace's diffusers stack, not Comfy's native loader - so it needs those packages installed even though they're not the usual Comfy deps.

Models come from the author's HuggingFace mirror (smthem/LTX-2.3-test-gguf) or Lightricks' own repos. The transformer GGUF goes in ComfyUI/models/gguf/, LoRAs in ComfyUI/models/lora/.

Troubleshooting

• "It loads, then OOMs." The model fits 6GB VRAM, but the peak system RAM is ~48GB (less if you enable virtual memory). If you're at 32GB RAM, budget for the swap file. RAM, not VRAM, is the real gating resource on 2.3.
• Everything is slow after a generation. Every node in this pack calls clear_comfyui_cache() on execute, which unpatchs and flushes loaded models. You pay a reload tax per run. That's by design for low-VRAM safety, not a bug.
• Nothing matches the example workflow. Check your quant/checkpoint names against the README's model tree exactly - this pack's combo boxes only list files that exist, so a missing file simply won't appear.