LTX 2.3 Spicy · Image to Video

Why 2.3 specifically

If you read one thing about LTX before using it: 2.3 is the version that turned the family around. Through the 19B era the model had a real reputation problem - image-to-video that froze into a slow zoom instead of moving, soft fine detail, prompt drift on anything complicated. Lightricks' March 2026 announcement led with that failure list rather than a feature list, which is why it landed so well: rebuilt VAE and a new latent space, a bigger text connector, reworked training specifically to kill static I2V, a new vocoder, and the family's first natively-trained portrait mode. The community confirmation came within a day and the line that stuck was that it had become the best video model you can run locally, considered as a whole.

So when you see two LTX image-to-video nodes in this pack with wildly different prices - the 19B at about a cent and a half per second, this one at about two - this is where the difference is. Spicy here means the uncensored build, which matters because the LTX-2 licence's acceptable use policy bans sexually explicit output while the weights themselves carry no filter at all. A hosted application deciding it will run that is a decision about its own terms, not a claim about Lightricks'.

How it runs, and the inputs

Stills go up, MP4 comes down, decoded into a normal ComfyUI VIDEO for Save Video. Your opening frame is re-encoded to PNG - JPEG as a fallback under the 10 MiB upload ceiling - and cached per account for twelve hours, so a rerun with a different prompt doesn't re-upload the same picture.

The fields:

• preset - the one input unique to the Spicy builds. tuned is the recommended general-purpose style stack; original applies light styling only. That's the author's description, and the practical read is that original is closer to the bare model: if your prompt keeps fighting the look, try original before you rewrite the prompt.
• image_url - the opening frame. Content lives in the image here; the tooltip on the sibling nodes says it plainly, the frame carries the content and the prompt is optional.
• prompt - the action, the camera, the mood. Optional in the sense that the frame does the heavy lifting, not in the sense that it doesn't help.
• duration_seconds - 3 to 20, whole seconds.
• resolution - 480p by default, up to 1080p. Start low; the defaults on this node are chosen for cost, and 2.3's detail improvements are exactly what you'd be paying more to see.
• seed - a real one. Same seed and settings, same clip.

Outputs are video and task id, the latter being the string that names the exact request on SpicyAPI if you need to ask support what happened.

There's no audio field on this node. If you want LTX generating sound with the picture - its actual signature - you want one of the other LTX nodes in this pack, not this one.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Restart ComfyUI, then Settings > SpicyAPI and paste a key from spicyapi.ai/console/keys; restart once so the node list is built from your account. On a server, SPICY_API_KEY in the environment instead - the key stays in the ComfyUI user folder and never lands in a workflow or a saved image's metadata. Nothing else installs: no dependencies, no weights, no GPU. Manager users can search the pack title, SpicyAPI. Recent ComfyUI (V3 node API, tested on 0.39) and Python 3.10+.

Worth knowing

You can't tell here whether the seed you liked will behave. LTX 2.3's prompt adherence was still its weakest axis at release, and the workflow that the community converged on was volume: many short candidates on different seeds, pick the winner. At roughly two cents a second, a 5-second 480p test is about ten cents, so that habit is affordable - do that rather than iterating wording for twenty minutes.

preset interacts with your prompt. The tuned bundle is a style stack; asking a heavily-styled model for a flat, unglamorous look is a fight. original is the escape hatch.

Billing mechanics are the same across this pack. Seeds start on fixed, so pressing Queue again with an unchanged graph reuses the cached result instead of buying a second clip; switch control after generate to randomize when you want a new draw. Max cost per run in the settings panel refuses anything quoted above your ceiling before a charge is made. And cancelling stops your wait, not the job - the task keeps running on SpicyAPI, is charged if it succeeds, and its output stays in your console for up to 14 days.