LTX Sequencer

LTX Sequencer is where the WhatDreamsCost pack started getting serious. Before LTX Director came along and ate its lunch, the Sequencer was the go-to way to build first-frame/last-frame (FFLF) videos and multi-keyframe shot sequences in LTX - and it's still the node the author points to when he says "use this instead of LTX Keyframer." Think of it as Keyframer plus the parts that were missing: it injects your guide images into the latent and appends the matching keyframe info to your conditioning, so you're not wiring two nodes to do one job.

It's an overhaul of ComfyUI's stock LTXVAddGuideMulti node, which tells you both what it does and where it sits in the LTX world. LTX is a first/middle/last-frame conditioning model - you anchor certain frames with images and it hallucinates the motion between them. The Sequencer is the cleanest stock-adjacent way to do that with any number of anchors.

How it works

Mechanically it's a subclass of ComfyUI's LTXVAddGuide, and the source reads like a well-behaved version of it: for each image in your batch it VAE-encodes the frame, works out which latent frame it lands on (converting pixel-space positions through the VAE's scale factors), and calls append_keyframe - which writes the encoded guide into the latent, updates the noise mask, and appends guide info to both positive and negative conditioning. Images beyond the batch size are skipped cleanly.

The thing it adds over Keyframer is an insert_mode. Set it to frames and each image gets an insert_frame_N (pixel space, negatives index from the end). Set it to seconds and you get insert_second_N instead, converted to frames using the frame_rate input - which is what you want when you're thinking "image B should land two seconds in," not "at frame 48." A num_images counter (up to 50) decides how many of those per-image widget pairs you actually see.

There's also a sync feature: multiple Sequencer nodes on a canvas can be synced so editing one updates all of them in real time - genuinely useful when you're iterating a shot list.

Inputs that matter

• positive / negative / vae / latent - the conditioning pair, the video VAE, and the latent to extend.
• multi_input - batched images, typically from the pack's Multi Image Loader, which auto-populates the widget list.
• insert_mode - frames or seconds. This is the toggle to learn.
• Per image: insert_frame_N / insert_second_N and strength_N (0–1).

Outputs

positive, negative, and latent - the latent has the guides baked in and a matching noise mask, and the conditioning carries the keyframe info. All three go into the sampler, with the latent coming from whatever empty-video-latent node you're using.

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, no extra deps - it leans on ComfyUI's own LTX extras (comfy_extras.nodes_lt). Pair it with current ComfyUI-LTXVideo nodes for LTX 2.3.

Gotchas

The one failure mode you'll actually hit: "Conditioning frames exceed the length of the latent sequence." That's an assert from the underlying LTXVAddGuide, and it means one of your insert frames (or the encoded guide's length) overruns the latent you fed in. Make your empty latent long enough to cover the last keyframe. And if you're starting fresh with LTX 2.3, know that Director supersedes this node for new projects - Sequencer is the right tool when you want stock-node simplicity without a full timeline editor in the middle of your graph.