LTXV Rebase Guides (segment-safe)

If you've ever tried to make a long LTX video by chaining segments - start clip two from clip one's last frame, the whole FLF/FFLF thing - you've hit the wall this node exists for. LTX is a first-frame/last-frame model with time-windowed conditioning. You set guides (a start image, an end image, maybe a mid text prompt) and each one carries a keyframe_idxs tensor telling the sampler which frames in the original timeline it applies to. Chain a second segment and those indices are suddenly wrong: they point at absolute frames that don't exist in your new clip's shorter latent, or they drag the previous segment's guides into frames where they don't belong. Enter LTXV Rebase Guides (segment-safe) - a tiny node that rewrites that guide metadata so carried-over conditioning still makes sense in a new segment.

It's from ato321/ComfyUI-LTXGuideRebase, a one-node, one-commit pack. Don't let the zero install footprint fool you: it doesn't touch latents, doesn't load a model, doesn't call any API. It's pure conditioning surgery, which makes it about as safe a custom node as this ecosystem has. Here's what it actually does.

How it works

LTX conditioning isn't just the prompt embedding. When you use Comfy-Org's LTXVAddGuide (or any guide node) to pin a frame or a text prompt to a window of the clip, it stuffs a keyframe_idxs tensor into the conditioning dict - the same metadata LTXVCropGuides reads. The tensor's shape is [batch, 1, entries, coords], where the coords hold start and end frame positions.

This node reads that field, does one of a few transforms, and writes it back via ComfyUI's own node_helpers.conditioning_set_values. "Rebase" means subtracting segment_start_frame from every index, so a guide that pointed at global frame 97 now points at local frame 0 in your new segment. "Drop outside" means keeping only guides whose window overlaps [segment_start_frame, segment_start_frame + segment_length_frames), so segment one's closing-shot guide doesn't ride along into segment two. It also mirrors keyframes between the positive and negative sides when only one of them has them, keeping the pair aligned.

The inputs that actually matter

• positive / negative - your conditioning. If neither carries keyframe_idxs, the node is a glorified pass-through; it only does work when guide nodes have been upstream.
• segment_start_frame - where this segment begins in the original/global timeline. This is the offset everything gets rebased by.
• segment_length_frames - the window for filtering. Default -1 means no filtering, rebase only. Set it to the segment's length and guides falling outside get dropped. Easy to forget this one defaults to off.
• mode - rebase_and_drop_outside (default), drop_outside, rebase_only, and clear_all, which strips all guides by setting keyframe_idxs to None. That last one is handy when you're starting a fresh clip that should not inherit the previous segment's guide state.

Outputs are the updated positive and negative conditionings, which feed straight into your sampler just like the originals.

Installing it

ComfyUI Manager: search "LTXVGuideRebase" and install. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/ato321/ComfyUI-LTXGuideRebase

Restart ComfyUI. No pip packages beyond what ComfyUI already ships (torch, node_helpers), no model files, no VRAM cost. The README also suggests just copying or symlinking the folder into custom_nodes if you want to hack on it.

When to reach for it

This is glue for segment-chaining workflows, not a daily driver. You want it in the extend/rebase portion of an FFLF chain where you carry conditioning forward and need the guide indices to stay in bounds - which the community has found fiddly with the stock AddGuide nodes. The "segment-safe" in the name is the whole pitch: rebase without risking out-of-range keyframes breaking the sample. Worth knowing the failure modes are quiet ones. If the tensor shape is unexpected the node returns it unchanged (safe, but silent). If window filtering drops every guide, keyframe_idxs becomes None - guides silently off for that segment. And if your conditioning has no keyframes at all, nothing happens, which surprises people who drop it into a plain T2V graph expecting a visible effect. If that's you, you don't need this node yet - you need guides first.