LTX Keyframer

LTX Keyframer is the pack's overhaul of the older LTXVImgToVideoInplaceKJ node, and it exists to solve one thing: getting your images into the right frames of a video latent without doing manual math. LTX video is a first-frame/last-frame conditioning model at heart - you tell it where a still must appear and it fills in the motion around it. This node is how you place those stills.

It belongs to the WhatDreamsCost-ComfyUI collection, the same free LTX-focused pack that grew into LTX Director. And here's the honest part before we go any further: the author recommends you use LTX Sequencer instead of this node. His own words - "after further testing it seems superior in at pretty much everything." Keyframer is still in the pack because people were already using it and it does one job slightly differently. Knowing that up front saves you a migration later.

How it works

The mechanism is straightforward and worth understanding because it's the same idea that powers the bigger LTX nodes: you feed it a video latent, a batch of images, and a set of (frame, strength) pairs, and it VAE-encodes each image and writes it into the latent at the specified position. Where it differs from the Sequencer is how it inserts: Keyframer replaces the latent frames at that location with the encoded image, and marks those frames in the noise mask with 1 - strength. Frames at strength 1.0 are frozen solid - the model regenerates nothing there, which is exactly what you want for a hard first/last frame. Lower the strength and the model gets room to reinterpret.

A couple of details in the source are worth knowing:

• Negative frame values work. insert_frame is in pixel space, and negatives index from the end of the video - so -1 reliably means "the very last frame," which is the classic last-frame trick without computing frame counts.
• It upscales your images to match the latent's resolution (bilinear) before encoding, so mismatched input sizes don't crash it.
• num_images dynamically configures how many insert_frame_* / strength_* widget pairs you see - up to 50. You don't have to scroll through all of them.

The inputs that matter

• vae - the video VAE (LTX's), used to encode your images.
• latent - the video latent you're inserting into.
• multi_input - the batched images, typically from the pack's own Multi Image Loader, which feeds every image down one wire and auto-updates the widget count.
• num_images - how many insert slots to configure.
• For each image: insert_frame_N (pixel-space position, negatives = from the end) and strength_N (0–1, how frozen that frame is).

Outputs

Just one: latent, with the images baked in and the noise mask updated. That goes straight into the sampler. The conditioning side - telling the model what each keyframe is - is handled by pairing this with a guide/conditioning node, which is exactly why the Sequencer's extra outputs (positive/negative conditioning) make it the better all-round choice.

Install

Standard for the pack:

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

Or search WhatDreamsCost in ComfyUI Manager, install, restart. No model downloads; it uses ComfyUI's own VAE and tensor utilities. Keep ComfyUI-LTXVideo current if you're pairing it with LTX 2.3 workflows.

Gotchas

The insertion model (replace + mask) means ordering matters: if two images land on the same frame, the later one wins. And remember it's frame positions in pixel space - if your latent was downscaled, the frame index you type needs to be the pixel-space frame (multiplied by the VAE's temporal scale factor), which is exactly the kind of arithmetic the Sequencer and Director were built to spare you. Use those, honestly.