LTX 2.5 · Text to Video

What it is

Prompt in, up to 20 seconds of continuous video out, in 16:9 or 9:16, at 720p through 4K, with a generated soundtrack included on every clip. No image needed. This is the hosted version of the LTX generation Lightricks pre-announced in January 2026 as their architectural jump - a new latent space aimed at preserving spatial and temporal detail, rather than another pass over the existing one. The 2.3 release before it is the one that fixed the family's motion reputation; 2.5 is where "long continuous take with sound" is the pitch rather than a party trick.

Worth knowing what you're buying in the family's own history, because it reframes the local option: LTX earned a following on speed and length, and kept the audio. What it never had was a clean install. The LTX-2 line ships a 22 GB Gemma text encoder, quantisation workarounds, and (from 2.3) resolution rules that fail instead of rounding - width and height divisible by 32, frame counts divisible by 8 plus 1. Running it locally is a project. This node is the same family's newest latent space, at from $0.09 per second, with no project attached.

Inputs

Four fields, and that's the whole node. There's no seed-less ambiguity here - you get a genuine seed - and no image, reference, mask or LoRA socket to configure.

• prompt - "describe the picture and the action you want." That's it, straight from the node. Every clip comes with generated sound whether or not you describe it, so if the audio matters, say so in the prompt; if it doesn't, expect the model to make something plausible anyway.
• resolution - 720p default, with 1080p, 2K and 4K above it. Each step up raises the per-second rate. The node's badge shows the current rate for the options you've picked, so you can see the cost move before you queue.
• aspect_ratio - 16:9 or 9:16.
• duration_seconds - 5 to 20 whole seconds. Continuous: this is one shot, not a multi-cut edit, so a prompt with three scene changes in it is asking for trouble.

Outputs are video, an ordinary ComfyUI VIDEO for Save Video or any downstream node that takes one, and task id - a plain string naming the exact request on SpicyAPI. Keep it; it's the fastest thing to hand support if a run misbehaves.

Install

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

Restart ComfyUI. Then Settings > SpicyAPI → paste a key from spicyapi.ai/console/keys → restart once so the node list is built from your account instead of the release snapshot. On a remote server, export SPICY_API_KEY instead; the key is stored outside your workflows, in the ComfyUI user folder, so it can't leak through a shared graph or a PNG's metadata. Nothing else to install: no dependencies, no model files, no GPU. Manager users can find it by searching the pack title, SpicyAPI. Recent ComfyUI (V3 node API, tested on 0.39), Python 3.10+.

Honest notes

Nine cents a second is the expensive end of this pack. A 20-second 4K clip is real money, and text-to-video has no cheap draft input to anchor it - you're paying to find out whether the prompt worked. The other LTX nodes here cost a fraction of this. If you're still writing the prompt, write it somewhere cheaper first, or lean on the fact that the seed is repeatable so a re-run at a lower resolution at least tells you something.

Prompt adherence is the family's weak axis, historically. LTX's own community record through 2026 is that Wan followed prompts better and LTX traded some of that for speed, length and audio. Generate several takes on different seeds and select, rather than grinding a prompt into shape - the seed input makes that a sane strategy.

Multi-shot prompts waste the model. This is one continuous take. If your idea needs cuts, that's the multi-shot options on the Kling and Wan nodes, not this one.

The licence bans explicit output in its acceptable use policy while the weights carry no such filter. LTX-2's community licence covers releases from January 2026 onward, is free below $10M revenue, allows hosting outright, and is revocable on notice with derivatives required to be share-alike. Note the gap between what the model will do and what the terms permit - and that a hosted endpoint layers its own policy on top of both.

The usual billing mechanics. Seeds default to fixed so re-queueing an unchanged graph reuses the cached clip rather than buying a new one; Max cost per run in the settings panel rejects anything quoted above your ceiling before it's charged; and cancelling stops your wait, not the job - it finishes on SpicyAPI, is charged if it succeeds, and lives in your console for up to 14 days.