LTX-2 Context ? Latent (continue) ??

The author's docstring calls this "IAMCCS-native equivalent of TTP's LTXVContext_TTP" - and if you've done LTX long-video work, you know exactly what that means. It's the node that takes the tail of your previous video, VAE-encodes it, and injects those frames as locked context at the start of your next segment's latent, with a noise mask so the sampler treats them as ground truth. Everything in the previous segment stays visually continuous because the next segment literally begins inside it.

Where the other overlap node in this pack works at the latent level (you bring your own prev_latents), this one works at the pixel level: you hand it the previous video as an IMAGE batch, and it does the encoding for you. That's the version you reach for when your pipeline has decoded frames lying around.

The inputs that matter

• previous_video - the previous segment's frames as an IMAGE batch.
• vae - required here (not optional), because the node has to encode.
• latent - the empty latent for the next segment; the context gets embedded into its front.
• enable - master switch. false = pure passthrough, context injection disabled. Handy for A/B testing continuity vs. independence.
• context_latent_frames (default 6) - how many latent frames to embed at the start. The tooltip flags the mapping: "LTX uses 8n+1 mapping," meaning a given number of latent frames maps to a specific pixel count. 6 is a sane starting point.
• exclude_last_frame (default true) - drops the very last frame of previous_video when building context, which the author says "often reduces over-constraint." Leaving it on is the smart default; the very last frame is usually the most degenerate.
• context_strength (optional, default 1.0) - 1.0 = fully locked context, 0.0 = no lock. Dialing this below 1 is your seam-tuning knob.

Outputs: latent (next segment's latent with context embedded) and report (what was injected).

How it fits the big picture

This is the "context" flavor of LTX-2 continuation, one of three complementary approaches in the pack: context injection (here), direct overlap copying + locking (Condition Next Latent), and segment bridge blending (Blend Latent Bridge). They all solve the same seam problem with different trade-offs. Context encoding costs a VAE pass but gives the model the most honest picture of what came before - it's the strongest continuity of the three, at the price of the extra encode.

Install & requirements

Standard IAMCCS-nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

or ComfyUI Manager → search IAMCCS, restart. No extra deps - but you need a real LTX-2 VAE wired in and LTX-2 loaded in ComfyUI, since without the native LTX-2 stack there's nothing to continue.

The trap

Over-constraint is the classic failure: too many locked context frames and the new segment recycles the old motion instead of progressing. If your continuation looks like a loop, shorten context_latent_frames and/or lower context_strength. And keep exclude_last_frame on unless you have a specific reason - the author added it because the last frame over-constrains.