🅛🅣🅧 LTXQ8Patch

LTXQ8Patch is how you turn on LTX's Q8 (8-bit) quantized attention and feed-forward kernels. Patch your model with it and the heavy matrix math runs in FP8/INT8 instead of full precision, which on the right hardware means faster generation and a smaller memory footprint. On LTX it's been a real speed lever since the 13B days - and, historically, a real source of installation pain. More on that below.

Quantization here is about the compute, not just storing weights small. The node swaps in kernels that do the attention and FFN operations in 8-bit. The catch, straight from the community record: without native acceleration on your GPU, Q8 mostly just "helps squeeze into memory constraints" rather than actually running faster. So whether this is a speedup or merely a VRAM saver depends entirely on your card.

How it works

You hand it a MODEL and it returns a patched MODEL with the selected parts quantized. You choose a quantization_preset to match your checkpoint, then toggle which sub-layers get quantized - self-attention, cross-attention, feed-forward - plus an optional FP8 attention path.

The inputs and outputs that matter

• model (MODEL) - the LTX model to patch.
• quantization_preset (enum: 0.9.8, ltxv2, full_bf16, custom) - the big one. Pick the preset that matches your model version: 0.9.8 for the older 2B-era checkpoint, ltxv2 for the LTX-2 family, full_bf16 to run unquantized bf16, custom to drive the per-layer toggles yourself. Choosing the wrong preset for your checkpoint is the most common way to get garbage or errors.
• use_fp8_attention (BOOLEAN, default false) - route attention through an FP8 path. Helps on hardware with native FP8 support (Ada/Blackwell), less so elsewhere.
• quantize_self_attn / quantize_cross_attn / quantize_ffn (BOOLEAN, all default true) - which parts to quantize. The defaults quantize everything; back individual ones off if a specific layer is hurting quality.

Output is a patched MODEL for your sampler.

How to install it

The pack installs through ComfyUI Manager (Ctrl+M → Install Custom Nodes → search LTXVideo → Install → restart), which the README recommends. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Lightricks/ComfyUI-LTXVideo

then restart ComfyUI. The node is under the LTXV menu. The Q8 path may also want the LTX Q8 kernels installed on your system - and that's historically been the fiddly part (see below). The LTX model itself remains the heavy dependency: 32GB+ VRAM and 100GB+ disk per Lightricks.

Common issues & troubleshooting

It's not actually faster. This is the headline gotcha, and it's well documented. Q8 only speeds things up where your GPU natively accelerates the low-precision ops; otherwise it just trims memory. The Lightricks team said it plainly - "without native acceleration it's most likely not going to be faster, just help squeeze into memory constraints."

It flat-out doesn't work on your card. The Q8 kernels have not worked on RTX 30xx-series GPUs - multiple users confirmed this. If you're on Ampere (3060/3080/3090), don't expect the Q8 acceleration path; you're better off with a GGUF quant or plain fp8 weights for memory relief. Ada (40xx) and Blackwell (50xx) are where FP8/Q8 pays off.

The install fought you. You're in good company - the Q8 kernels were called "the most confuse installation of my life" and "instruction not clear for comfyui portable" back in the 13B era. If the kernels won't build or import, fall back to a community GGUF quantization, which sidesteps the kernel install entirely.

Wrong preset = broken output. Match quantization_preset to your checkpoint. ltxv2 for LTX-2/2.3, 0.9.8 for the old 2B line. A mismatch is the fast route to noise or a crash.