LTX2.5 Model (CRT AutoDL)

LTX2.5 Model (CRT AutoDL) is the centerpiece of CRT's LTX2.5 stack: a diffusion-model loader that offers four pre-quantized variants of Lightricks' LTX-2.5 distilled DiT and downloads whichever you pick on first run. It's the answer to the question that stops most people before they ever sample: "which checkpoint do I download, and will it fit my card?" This node has that built in - the dropdown names are literally the VRAM tiers.

LTX-2.5 is the follow-on generation to LTX-2.3, the one Lightricks flagged early as an architectural jump with a new, more detail-friendly latent space. The community quants this node pulls (from Joey Gambino's LTX-2.5-Quantized repo) are the distilled checkpoint repacked for consumer cards, and the four options map to how much memory you've got:

• 24gb (default) - the mix4x8 17 GB quant. The sensible default for the majority of people running 24 GB cards.
• 32gb - the int8 quant. The quality-per-byte winner if you've got the headroom; the highest-VRAM option here.
• 16gb 4x8mix - the 13.8 GB mix4x8 quant, for 16 GB cards.
• 16gb NVFP4 - NVFP4 (Blackwell 4-bit) at ~13 GB, the other 16 GB play.

Don't overthink it: if you have 24 GB, take the 24gb default. If you're on 16 GB, one of the two 16 GB options is the start, and you'll pair it with the Light CLIP and the offloading settings below.

The shared CRT fast settings

Like every AutoDL diffusion loader, you get the trio that controls how the model runs rather than what it is:

• attention_method - disabled by default, which is correct. The menu offers PyTorch attention plus a row of SageAttention modes (sageattn_qk_int8_pv_fp16_cuda, sageattn3, …) if you've installed SageAttention - these are the real speed/memory wins on Blackwell, and the reason the menu exists. Ignore until you've got the base workflow working.
• patch_cublaslinear - off by default; toggles ComfyUI's cublas_ops fast path.
• enable_fp16_accumulation - on by default, matching what the LTX community converged on for quantized runs.

Output: a single MODEL for your sampler. The rest of the LTX2.5 pipeline is the sibling AutoDL nodes - LTX2.5 CLIP (a w4a8 Light Gemma 4 12B encoder), LTX2.5 VIDEO VAE, and LTX2.5 AUDIO VAE for the synchronized-audio side of things. All self-downloading.

Installing it

Part of CRT-Nodes, installed once:

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

or ComfyUI Manager → search CRT-Nodes, restart. The pack's requirements are heavy (opencv, transformers, faster-whisper, librosa…), so the install is slow. The model download is the bigger wait - expect 13–17 GB depending on the quant, with a progress bar in the console.

Common issues

• First run "hangs" - it's downloading 13+ GB. Leave it; the console shows progress. Subsequent runs load from disk.
• OOM - you picked a quant too big for your card. Drop a tier, use the 16 GB options, and/or switch the CLIP to the w4a8 Light one. System RAM of 64 GB is the quiet requirement behind "16 GB VRAM" claims in the LTX world.
• Don't run the stock dev checkpoint with these settings - these are distilled-model quants. If you swap in a non-distilled checkpoint, the fast settings and step counts that work here won't transfer.
• Update breakage - red node or NaN → right-click → Fix node (recreate).