LTXVCropGuides

If LTXVAddGuide is the node that lets you place keyframes inside an LTX video, this is the node that keeps those keyframes from showing up in the finished clip. When LTX's guide mechanism injects a first frame, a last frame, or a mid-video fragment, the guide latents get appended to the latent sequence and their positions get recorded in the conditioning. LTXVCropGuides is the cleanup pass: it reads how many guide keyframes are sitting on the conditioning, trims them off the end of the latent, clears the keyframe bookkeeping, and hands back a latent that's exactly the video you asked for. It's the boring node that makes the fancy one usable, and forgetting it is a classic.

How it works

The node takes your positive, negative, and latent and inspects the positive conditioning for keyframe_idxs. If there are no guides, it passes everything through untouched - it's a no-op in the common case, which is why you can leave it in a graph permanently. If guides exist, it counts them, trims that many latent frames off the end of the tensor (and the matching slice off the noise mask), and strips the keyframe_idxs and guide_attention_entries from both conditionings. Three outputs - positive, negative, latent - clean and ready.

The why: LTX guides are implemented by appending the encoded guide image/video onto the latent sequence in the time dimension. That means a first+last-frame (FFLF) setup leaves extra latent frames hanging off the end of your clip. If you decode without cropping, you get a video that's the wrong length or shows the keyframe material bleeding into the tail. Cropping removes exactly the bookkeeping and the appended frames.

Where it fits in the workflow

The standard chain looks like: LTXVAddGuide (frame 0) → LTXVAddGuide (last frame, negative index) → LTXVCropGuides → sampler → VAE decode. You'll also see it between successive AddGuide calls in complex keyframe sequences, and it's a permanent resident of the community's FFLF and shot-sequencing templates.

How you get it

Ships with ComfyUI core - no install. No model files of its own; it only manipulates conditioning and latent tensors, so it works with any LTX checkpoint and VAE you already have in models/.

Where people get burned

Forgetting it entirely. Skip the crop and your clip comes back longer than requested, with the keyframe frames visible at the end - often mistaken for "LTX extended my video weirdly." Double-cropping. Run it twice in a row with no new guides in between and the second pass is a harmless no-op, so don't panic about that. Version drift with 2.3. The rebuilt VAE and latent space in LTX-2.3 changed how guides encode, so if a workflow built for LTX-2 produces a jumpy or over-long clip on 2.3, the guide+crop pair is the first thing to verify rather than the sampler settings. The pattern "AddGuide, then CropGuides" is the same across all of it - the arithmetic just has to match the checkpoint's compression.