LTX 2.5 · Image to Video

The top of the LTX line

LTX 2.5 is the checkpoint Lightricks pre-announced in their January 2026 AMA as the architectural jump - a new latent space, better preservation of spatial and temporal detail - as distinct from the incremental 2.1/2.3 work on the same architecture. The roadmap named 2.1 and 2.5 specifically and said yes to the obvious question about open weights. Along the way the company's LTX arm spun out into an independent business in July 2026 with the open-weights commitment restated, so 2.5 is the same house with a new letterhead.

Here it is as a hosted node: 5 to 20 seconds of continuous video from one still, at 720p, 1080p, 2K or 4K, with a generated soundtrack on every clip and a closing frame you can pin. From $0.09 per second, which makes it the expensive end of this pack - about six times the LTX-2 19B rate. That's the trade: the newest latent space and a 4K tier, versus fourteen-tenths of a cent on the older 19B.

Inputs

• image_url - the opening frame, and it decides the shape of the clip. The node's own tooltip asks for an "authorised" frame, which is the platform's boilerplate, not a subtle hint that you must own the rights to your own photo.
• prompt - the action and the camera, and nothing else, per the tooltip. This is 2.3-and-later LTX, where the motion problem was specifically reworked; give it a clear movement rather than a paragraph of mood.
• resolution - 720p by default, with 1080p, 2K and 4K above it. Higher tiers cost more per second, and 4K at 20 seconds is a serious sum; the node's badge shows the live rate before you commit.
• duration_seconds - 5 to 20 whole seconds, one continuous shot. Long single takes are LTX's differentiator, and 2.5 is where that's most worth paying for.
• last_image_url - optional, and a genuinely nice feature: the closing frame the clip should resolve to. Pin a specific end state and LTX interpolates the motion between your two images, which is how you close a loop or land on a key visual instead of hoping.
• seed - real, repeatable. Same seed and settings, same clip.

Outputs are video, a standard ComfyUI VIDEO that Save Video writes, and task id, the string identifying this particular request on SpicyAPI.

Install

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

Restart ComfyUI, open Settings > SpicyAPI, paste a key from spicyapi.ai/console/keys, and restart once more so your node list is rebuilt from your account rather than the bundled snapshot. Servers can set SPICY_API_KEY in the environment; the key is written to the ComfyUI user folder, never into a workflow file or an image's metadata, so sharing graphs and pictures is safe. ComfyUI Manager users can search the pack title, SpicyAPI. Nothing else to install - no pip step, no weights, no GPU. Current ComfyUI (the pack uses the V3 node API and is tested on 0.39), Python 3.10+.

The catches

This is the priciest video node in the pack. Nine cents a second means a 20-second 720p clip is about $1.80, and 4K pushes the per-second rate higher still. The cheapest way to use it well is at 720p for tests and up to 4K only for the take you've already chosen - which the seed input lets you do, since the same seed and settings reproduce the same clip. Set Max cost per run in the settings panel and an over-budget quote stops before any charge.

The upload ceiling is 10 MiB per image. A high-resolution frame is re-encoded to PNG and falls back to JPEG, and only then refuses with a "scale it down first" error. Very large photographs need pre-shrinking, which is a small annoyance given you're about to ask for 4K output.

Seeds start on fixed, deliberately. Every run costs money, so a second Queue press with nothing changed reuses the first result instead of quietly buying another clip. Change control after generate to randomize when you actually want a fresh draw.

Cancelling stops the waiting, not the job. The task keeps running on SpicyAPI, is charged if it succeeds, and stays in your console for up to 14 days. If you bailed because the estimate looked steep, it may still be steep.

Licence, plainly. LTX-2's community licence - free below $10M revenue, hosting allowed, derivatives share-alike, revocable on notice - self-scopes to LTX-2 releases from January 5, 2026 onward, and its referenced acceptable use policy bans sexually explicit output even though the weights carry no filter. That's the terms of the model, not of this node; the hosted endpoint applies its own policy on top. If you need something the platform's filter refuses, hosted is the wrong architecture - no node can patch a refusal out of a server.