LTXV Context Windows

LTXV Context Windows is the LTX-family flavor of ComfyUI's context-window sampling, and LTX is a great candidate for it because LTX is the fast one. Lightricks' video model has been the speed king of local generation since the 0.9.x era - 121 frames in a handful of seconds on a strong card, running on far less VRAM than Wan or Hunyuan. Its weakness was never speed; it was that long clips strain the temporal attention the same way every video model does. This node slides a window over the frames during sampling so you can push past LTX's comfortable clip length without the memory bill climbing linearly with every frame.

The node's units are LTX-native: context length is in real frames and must be 8n+1, matching LTX's temporal compression. That's the tell that this wrapper is doing the conversion math for you - the generic manual node makes you think in latent positions, and this one doesn't.

How it works

Same family machinery as the Wan and manual context nodes: install a context handler on a cloned model, slice the frame sequence into overlapping windows, sample each window's attention, fuse the results. The LTX wrapper pins the windowing to the temporal dimension (dim 2), defaults freenoise on (noise shuffling that stops window boundaries from showing), and converts your real-frame values into latent values ((length-1)//8 + 1).

One LTX-specific detail: because LTX-style I2V carries the reference image in the latent rather than in separate conditioning channels, the node sets both cond_retain_index_list and latent_retain_index_list to '0' when you enable retain_first_frame. In plain terms: turn that on for image-to-video and the start frame stays present in every window instead of being forgotten after window one.

The inputs that matter

• context_length - default 145 real frames, must be 8n+1. That's the kind of length this node unlocks; the 0.9.x models were comfortable in the 81-161 frame range, and windows around 145 keep attention cheap.
• context_overlap - default 40 real frames. The overlap is your seam-control knob.
• retain_first_frame - off by default; for I2V, turn it on so every window keeps the first-frame reference.
• context_schedule - default standard_uniform; looped_uniform + closed_loop if you want seamless loops. Leave the defaults to start.

Output: the patched MODEL for the sampler. Ships inside ComfyUI core.

Common issues & troubleshooting

Motion discontinuities at window boundaries. With LTX's aggressive temporal compression, windows are short in real time, and cutting between them can read as a hitch. Raise context_overlap first; if it persists, try a different fuse_method (the pyramid default weights each window's center most).

The fast-distilled models want fewer, not more, steps per window. LTX distilled checkpoints run at 8 steps and CFG 1. Context windowing doesn't change the step recipe - if your windowed clip looks worse, the model is being over-sampled or under-sampled per window relative to its training, not the windowing's fault. Keep the distilled recipe identical.

I2V reference drift across windows. Enable retain_first_frame. If the drift survives, bump the overlap so the reference window's influence reaches further into the next one.