🅛🅣🅧 LTXV Img To Video Condition Only

Image-to-video is the mode LTX is actually good at - the community consensus is that you get far better results animating a solid still than generating video from text cold. This node is one of the pieces that makes I2V work: it takes an input image and locks the first frames of your video onto it, so the clip starts from the picture you chose instead of whatever the model dreams up.

The name to pay attention to is "Condition Only." There's a full I2V node that builds an empty latent and applies the image in one move; this one skips the building and only does the conditioning, onto a latent you already have. In the node's own words, it "applies image conditioning to the first frames of an existing latent" and "creates a noise mask to control conditioning strength." That's a building-block, not a one-click node - you reach for it when you're constructing the latent yourself: video extension, keyframe setups, two-stage pipelines, anywhere the latent already exists and you just need to stamp your starting image into it.

How it works

Under the hood it VAE-encodes your image into LTX's latent space and writes it into the opening frames of the latent you passed in. Then it builds a noise mask - a per-region instruction to the sampler that says these frames are anchored, preserve them; the rest, denoise freely. That mask is the mechanism. When the KSampler runs, the conditioned frames stay close to your image and the model generates outward from them. It's the same idea as inpainting a mask, moved into the time axis of a video.

The inputs that matter

• image - your start frame. In practice you generate a clean still elsewhere (Flux and Z-Image are the popular choices) and animate it here; a good reference image is most of the battle in LTX I2V.
• latent - the existing video latent to condition. This is the "on an existing latent" part; the node writes into it rather than creating it.
• vae - needed to encode the image into the latent space. Use the VAE that matches your LTX checkpoint, not a random one.
• strength (0–1, default 1) - the dial you'll actually tune. It sets how hard the first frames are pulled toward your image. And here's the trap: crank it to a full 1.0 and the opening is perfectly faithful but the clip often freezes - the slow-zoom, no-real-motion failure LTX I2V is notorious for. Back it off toward ~0.8 and you give the motion room to breathe while still respecting the source. Too low, though, and the video drifts off your image entirely. It's a balance, and 0.8 is the community's usual starting point.

There's also an optional bypass (default false) - flip it on to pass the latent through untouched, which is a clean way to A/B your clip with and without the image anchor without rewiring the graph.

The single output is latent - the conditioned latent, ready to wire into your sampler.

Installing it

It comes in Lightricks' ComfyUI-LTXVideo pack. Via ComfyUI Manager: Install Custom Nodes, search "LTXVideo", install, restart. By hand:

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

then restart; the node shows up under the "LTXVideo" menu. Pack-level note: LTX-2/2.3 is heavy (32GB+ VRAM, 100GB+ disk per the README), and the checkpoint plus the Gemma 3 text encoder download on first use. Start from one of the pack's example I2V workflows rather than wiring this from scratch - it's a component meant to sit inside a graph, and the examples show you where.

When it goes wrong

Frozen clip, no motion? That's the number-one LTX I2V complaint. Lower strength first; also try a higher frame rate (48fps is a known trick for reducing LTX's glitchy or static motion in I2V). Drifting away from your image? Raise strength. Colors or structure look wrong on the first frame? Check you're feeding the VAE that matches your checkpoint. One more, easy to miss: LTX nodes and workflows are version-sensitive - the 1.x, LTX-2, and LTX-2.3 lines use different latent spaces and Lightricks confusingly labels a lot of 2.3 material as "LTX-2." Make sure the workflow, the checkpoint, and this pack are all on the same LTX version before you debug anything cleverer. Worth knowing: I2V was LTX-2's weakest mode at launch and got a major overhaul in 2.3, so a lot of old "LTX I2V is broken" advice no longer applies.