LTX2 Efficient Video Sampler

LTX-2 is fast for what it is - Lightricks' 19B model that does synchronized video and audio - but on a consumer GPU it has a dirty secret: the attention cost is quadratic in frames × spatial tokens, so your card pegs at 100% and your temps climb while it "runs." This node is a drop-in sampler replacement that fights back with three ideas: freeze the parts of attention that stop mattering, throttle against a target temperature, and optionally sample keyframes instead of every frame. It's the flagship of the kakachiex2/comfyui-ltx2-efficient pack, aimed squarely at people on 6GB cards like the author's own RTX 2060.

What it replaces and how it works

You wire it exactly like the KSampler in a normal LTX-2 pipeline: model, latent_video, positive, negative go in, a LATENT comes out, and you feed that into a VAE decode (this pack's LTX2TemporalVAEDecode pairs nicely). Under the hood it does a few things the stock sampler doesn't:

• Spatial attention freezing. LTX-2's transformer blocks have two attention paths: attn1 (spatiotemporal self-attention) and attn2 (cross-attention to your text). After freeze_ratio of your steps, the node skips attn1 entirely while always running attn2, so prompt adherence stays intact while the expensive part gets cheaper. Structure converges early in diffusion; the last steps mostly refine motion, which is the idea.
• Thermal throttling. It reads real GPU temperature through pynvml and inserts delays between steps to hold target_temp. The Thermal Auto-Scale preset (the default) adjusts the delay dynamically instead of using a fixed one.
• Built-in FlowMatch sigmas. LTX uses flow matching, not plain noise schedules. The node generates LTX-style sigmas internally (max_shift 2.05, base_shift 0.95, terminal 0.1) and injects frame_rate into the conditioning - or you can override it all by wiring an LTXVScheduler output into the optional sigmas input.

The inputs that actually matter

You don't need to touch most of the 20-odd knobs. The handful worth knowing:

• optimization_preset - pick your GPU tier: Quality - Fast (RTX 3080+) down to Quality - Ultra Cool, plus Thermal Auto-Scale (default) and Custom. The presets override the manual knobs, so don't fight them.
• target_temp - default 70°C; drop it toward 60-65 if your card still runs hot.
• freeze_ratio - how early attention freezing starts (0.3 default). Higher = faster and cheaper, lower = closer to full quality.
• optimization_engine - extra temporal-attention tricks: Adaptive Cache (default), Token Merge / VidToMe, Step Adaptive, or Combined. These are the "free speed" options; try the default before you go hunting.
• context_slice_method - LTX-2's dual text encoder (Gemma 3 12B + connector) produces 7680-wide context when the model wants 4096. Keep Last 4096 (Slot 2) is correct for Gemma+LTX Connector.

One honest gotcha before you get excited: the README's optimization table advertises presets with frame_stride 2-8. That's stale. The actual code hard-sets every preset to frame_stride=1 with a comment that says skipping frames breaks temporal coherence in LTX's full temporal attention - the efficiency comes from freezing and throttling, not from frame striding. The frame_stride + interpolation_method (slerp default) controls are still there, but treat frame_stride > 1 as a preview/experiment knob, not the main event. That's also exactly what the pack's separate experimental sampler is for.

Installation

The easiest path is ComfyUI Manager: search "LTX2 Efficient" and install. Manual install works too:

cd ComfyUI/custom_nodes/
git clone https://github.com/kakachiex2/comfyui-ltx2-efficient
cd comfyui-ltx2-efficient
pip install -r requirements.txt

Then restart ComfyUI. Heads-up: the README's manual-install block has a copy-paste slip - it clones a differently-named repo URL. Use the one above. The requirements.txt is just pynvml>=11.5.0, which is optional (it enables thermal auto-scaling; without it the node falls back to fixed throttling with a warning). No model files ship with this pack - you still load your LTX-2 checkpoint, Gemma text encoder, and VAE through ComfyUI-LTXVideo as usual.

Troubleshooting

• "pynvml not installed" warning - expected if you skipped the requirement. pip install pynvml (or nvidia-ml-py, pynvml is deprecated but works).
• Card still runs hot - lower target_temp, or check that your preset matches your actual GPU. A 3080 preset on a 2060 is asking for trouble.
• Flickery or broken motion - if you've set frame_stride > 1, that's the cause. Set it back to 1 and let the freeze ratio do the work.
• Quality issues - the presets in code keep freeze_ratio modest and frame_stride at 1, so if something looks off, first suspect your sampler/steps (LTX-2 dev models want ~20 steps, distilled wants ~8 at low CFG) before blaming the node.

It's a small, one-person pack with no big community footprint, so keep expectations realistic - but for a 6GB card trying to run a 19B model, the freezing-plus-throttling approach genuinely does what it says on the tin.