LTX2 Experimental Keyframe Sampler ⚠️

This is the rare node whose own documentation tells you not to expect much. LTX2ExperimentalKeyframeSampler is the "fine, you asked for it" version of the efficient sampler: it genuinely samples only keyframes (every Nth frame), interpolates the rest, and lets you crank the aggressiveness - while the node description and the module docstring both warn that frame striding breaks temporal coherence for LTX and "may produce distorted output." The pack's other sampler keeps frame_stride=1 in every preset for exactly this reason. So why does this node exist? Mostly for experiments, and for future models that don't use full temporal attention.

How it works

Mechanically it's a sampler with all the inputs of the main one - model, latent_video, positive, negative, seed, steps, cfg, sampler_name, denoise - plus a scheduler dropdown (simple, sgm_uniform, karras, exponential, ddim_uniform, beta, normal, linear_quadratic, kl_optimal) that the main sampler lacks. Then the experimental bits:

• experimental_preset - Conservative (stride=2), Moderate (stride=4), Aggressive (stride=8), or Custom. Each preset tunes freeze_ratio downward and throttle_delay_ms upward as stride climbs, which is the author acknowledging things get uglier the more aggressive you get.
• frame_stride - sample every Nth frame, then fill the gaps.
• interpolation_method - linear, slerp, or motion (the pack's optical-flow-ish warping). No none here, unlike the main sampler.
• attention_window, freeze_ratio, throttle_delay_ms - the familiar attention-freezing and pacing knobs, with throttle now a plain milliseconds value instead of temperature-driven.

Output is a LATENT. The interpolation code is the same interpolate() utility the main sampler uses - linear lerp, spherical slerp (better for latents since it preserves magnitude), and motion-compensated warping.

The honest verdict

The math is real - fewer diffused frames is a big attention win, which is why frame-stride tricks work in real-time video systems - but LTX-2 isn't built for it. Its temporal attention spans all frames, so skipping them produces warped, flickery, or just wrong motion. The experimental presets at stride 2 are borderline; at stride 8, per the source, you're "likely to break most models." If you're on a desperate VRAM budget, your realistic path is the main LTX2EfficientSampler with freeze ratio and thermal throttling, not this. Use this one if you're curious, prototyping, or testing a future model with local temporal attention.

Installation

Same pack as everything else here - kakachiex2/comfyui-ltx2-efficient. ComfyUI Manager (search "LTX2 Efficient") or:

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

Restart ComfyUI. No model files, no heavy deps.

Troubleshooting

• Distorted/noisy output - that's the documented behavior at frame_stride > 1. It's not a bug you can tune away; it's the fundamental limitation the node exists to explore.
• "Motion" interpolation is slow - expected; it's the heavy option. Try slerp for a fast, smoother default.
• Your real workflow needs reliability - switch to the non-experimental LTX2EfficientSampler. This is the node to play with when you have time to waste, not the one to build on.

It earns its ⚠️. Worth knowing about because it clarifies why the pack's main sampler is so conservative about frame striding - and it's a nice sandbox for people who want to see exactly what breaks.