LTXVImgToVideo

LTX's whole reason for existing is speed, and this is the node that gets you from a still image to a draft video in seconds instead of minutes. Where Wan takes twenty minutes a clip on a 4090, LTX 2B takes twenty seconds - the community standard practice became "draft in LTX, refine in Wan," and this node is the drafting machine. It's the natural I2V entry point: give it a good still, get a clip back fast enough that you can try ten prompts and keep the one that moves right.

How it works

The node takes your positive and negative conditioning, a vae, and an image, then does the classic first-frame trick: it encodes the image with the LTX VAE, writes those encoded frames into the front of an empty latent video tensor, and sets a noise mask over the conditioned frames so the sampler treats them as fixed facts rather than redrawing them. The strength input (0 to 1) is the dial on how hard the image is anchored: the mask over the image frames is 1 - strength, so strength 1.0 pins the first frame completely and strength 0.0 basically ignores it. The community's starting point for I2V image strength on the LTX-2 line is around 0.8.

The canvas inputs follow LTX's compression rules, and they matter: width and height step in multiples of 32 (they must be divisible by 32 on the newer checkpoints - off-grid values fail rather than round), and length must be 8n+1 - the default 97, and the other common values are 81 and 121. batch_size is there if you're doing parallel shots.

The output chain

Three outputs - positive, negative, latent. The latent goes into your KSampler with the LTX checkpoint; the whole I2V graph is: this node → sampler → VAE decode. ComfyUI's official LTX example workflows show it.

How you get it

Core node, no install. The models are the download: an LTX checkpoint (the 2B runs on 6GB VRAM, the 13B wants 16GB+ or GGUF quants, LTX-2 wants more and pairs with a Gemma text encoder), the matching LTX VAE, and the LTX text encoder, all in the standard models/ folders.

Where people get burned

Prompt quality is the skill. LTX is notoriously prompt-sensitive - short vague prompts underperform, and the standing advice is long, concrete, paragraph-length prompts describing motion. If you can't get the motion you want, that's usually the prompt, not the node. Frozen output. "No motion" was LTX-2's most complained-about failure, and it's worth knowing the 2.3 line fixed it - if you're on an older checkpoint, the fix is a better still and a more action-oriented prompt. Resolution games. LTX-2 renders internally at half your requested resolution and upscales, so to match Wan at 720p you ask for 1080p - quality expectations need recalibrating downward, then up. And the speed is the feature: run a handful of seeds and pick, because at this generation cost, seed-hunting beats prompt-polishing every time.