🎬 LTX Reference Sequence

LTX Reference Sequence is the multi-frame sibling of LTX Reference Conditioning. Same idea - encode a reference, prepend its tokens to the video sequence inside the transformer - but instead of one still image, it takes a frame sequence and uses a window of N frames as the reference. A single image can tell the model who someone is; a sequence can tell it how they move, how light plays across a scene over time, the rhythm of a performance. That's temporal style, and a still can't carry it.

It slots into the same architecture as its single-image sibling: requires LTX Reference Enable upstream to patch the model, and the reference tokens get the same 3D RoPE treatment so self-attention attends to them across all generated frames. If you've ever matched a moving subject to a reference clip and watched the identity hold but the motion feel wrong, this is the node that pulls the motion into the conditioning.

The inputs that matter

• model, vae, images - required. images is a multi-frame IMAGE (batch) - from a Load Video node, a frame sequence, or any upstream frame producer.
• start_frame (0–240, default 0) - where the reference window begins in the input sequence. Frames before it are discarded.
• num_frames (1–25, default 9) - how many frames from start_frame become the reference. The default 9 isn't arbitrary: LTX2's temporal compression is 8, and 1 + 8k gives you ~2 latent temporal positions. 17 and 25 are valid too - more temporal context, higher VAE cost.
• target_latent - same trick as the single-image node: wire in the LATENT going to your sampler and the frames are resized in pixel space to match before encoding.
• strength (0–2, default 1) - scales the reference latent; 0 bypasses and clears state.
• position_mode - reference (default) overlaps the target's first frames; prefix_continuous places memory before the target temporally.

Output is a MODEL for your sampler.

How it works

The window of frames is VAE-encoded as a batch and attached the same way the still-image node attaches its single latent - via model_options.transformer_options plus the diffusion-model attribute side-channel - and the Enable patches extend the per-frame compressed modulation tensors to match the prepended token count. The practical note from the tooltip: the default window is tuned to LTX2's temporal compression, so if you deviate from 1+8k frame counts you get whatever latent F dimension the VAE yields, which is fine but worth knowing before you assume 9 is "more is always better."

Installing it

Same TenStrip pack:

cd ComfyUI/custom_nodes
git clone https://github.com/TenStrip/10S-Comfy-nodes.git 10S_Nodes

or ComfyUI Manager → "10S". Restart, no extra dependencies.

Gotchas

Cost scales with window size - 25 reference frames is a lot of extra tokens in the attention sequence every step. And this is still LTX2-specific, so the same "wrong model → RuntimeError from Enable" rule applies. If your use case is pure identity with no motion context, the single-image node is cheaper and simpler; reach for Sequence when the reference clip's movement is part of what you want to preserve.