LTXVAddGuidesFromBatch

LTX Video has had a genuinely underused superpower since early on: you can inject reference frames - whole images, not just a start frame - into the middle of a generation, at whatever timestep you want, and the model treats them as guides rather than just conditioning the first frame. The catch is that Lightricks' own official example workflows never really showed this off, so for a long stretch the community had to discover the underlying AddGuide mechanism on its own and build wrappers around it before realizing how much control it actually offered. This node is that mechanism, batched: instead of wiring in one guide image at a time, you hand it an entire image batch, and it drops every non-black frame in as a guide at its matching position in the latent, in one pass.

What it's doing under the hood

Think of your images batch as a timeline with gaps. Anywhere you leave a solid black frame, nothing happens - that position just gets denoised normally. Anywhere you drop in a real image, that frame becomes a guide: a strong hint pulling the generation toward that content at that exact point in the sequence. It's the same idea as first-frame/last-frame conditioning, generalized to as many keyframes as you want, wherever you want them. This is the raw capability underneath higher-level tools like LTX Director's timeline editor - those just give you a UI for building the batch this node consumes.

The inputs and outputs that matter

• images - your guide batch. Real frames become guides; solid black frames are gaps. Get your frame indices right before you wire this in - a guide landing on the wrong frame is the most common way this goes sideways.
• strength (default 1, range 0–10) - how hard each guide pulls the generation toward it. Community reports on LTX-2's image-conditioning strength put a good starting point around 0.8, which is a reasonable anchor here too, though this node's guides can go well past 1 if you want the model locked hard to your reference.
• positive, negative, vae, latent - the standard conditioning/latent set you're already carrying through your LTX pipeline; this node reads all four and re-emits positive, negative, and latent with the guides baked in.

How to install it

It ships in KJNodes - ComfyUI Manager → search "KJNodes for ComfyUI" → install, or cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-KJNodes, then pip install -r ComfyUI-KJNodes/requirements.txt, then restart. This particular node is conditioning-only logic, so there's nothing extra to download for it specifically - but obviously you need an actual LTX Video checkpoint and its VAE already set up, since this slots into an existing LTXV pipeline rather than standing alone.

Common issues & troubleshooting

Frame alignment is everything. The node reads guide position from where the image sits in the batch, so your batch length and frame spacing need to match your target latent's frame count before you start tuning anything else. Get the plumbing right first.

"Non-black" means literally non-black. Pad your gap frames with actual pure black, not a near-black or slightly-off frame - anything with nonzero pixel values gets read as a guide, and an accidental guide at a frame you meant to leave alone shows up as a jarring discontinuity in the output.

Strength too high can fight the surrounding motion. Cranking strength up to force a guide to stick can produce a visible "snap" at that frame if the motion around it doesn't have anywhere to go. Back it off toward the 0.8-ish range as your default and only push higher when the guide isn't taking hold at all.