LTX 2.5 Can Now Edit Your Footage. Nine Tools Exist, and Seven Went Missing.
What you can change about a clip you already shot, which tools you lose by upgrading, and the five-minute check to run first.
Take a five-second clip. Open the first frame in Photoshop. Set the building on fire, or put a different jacket on the actor, or paint the grey untextured 3D model. Save it, hand it back with the original clip, and the change carries through the rest of the shot. Through camera cuts. With your motion, your timing and your framing untouched.
That is a free 654 MB add-on called LTX Ripple, published Sunday by one person working alone. It is the second such tool someone has released for LTX 2.5 in three days, and it arrived at the end of a week when Lightricks moved eight of its own editing tools onto the same base.
Nobody shipped a new video model in that window. The week's product was add-ons, and what they all take as input is footage that already exists.
Why small add-ons beat a better generator
For most of this year the thing worth waiting for in open video was a bigger, better model, and the question was whether the new one beat the old one.
That question has a ceiling. A better generator still starts from nothing, so every shot is an argument with a text box about a scene you have never seen.
These add-ons work the other way round. You hand one a video and it hands back the same video, changed. The composition is still yours. So is the camera move, and the performance.
The way they learn is worth thirty seconds of your attention, because it explains what they can and cannot do. During training the tool is shown pairs: an input and the finished result it should produce from that input. Show it enough pairs of black-and-white clip and coloured clip, and it learns colourisation. Show it grey clay render and textured render, and it learns texturing. At the other end you supply a new input of the same kind and get the matching result.
So the useful question stopped being "how good is LTX 2.5" and became "what kinds of input does it accept." Right now, nine.
What you can actually do today
Between September 8 and 10, Lightricks published eight add-ons for LTX 2.5, joining an upscaler from August. The shelf:
| Tool | You give it | You get back |
|---|---|---|
| Clean-Plate | your shot and a mask | the shot with that object gone |
| Day-To-Night | a daytime shot | the same shot at night |
| Colorization | black and white footage | colour |
| Deblur | soft footage | sharper footage |
| Decompression | compressed footage | the blocky artefacts cleaned up |
| Upscaler | small footage | bigger footage |
| Water-Simulation | a shot | water added to it |
| Ingredients | a reference sheet | a new clip using those characters and props |
| Cinemagraph | a still photo | one part of it moving, the rest frozen |
Download counts say people reach for the creative ones, not the repair ones. Ingredients leads the new batch by a wide margin.
Ingredients is worth understanding even if you skip the rest, because it fixes the thing that makes AI video useless for stories: your character looking like a different person in every shot. You build one composite image, a sort of contact sheet, with a clean face close-up, a body turnaround, each prop photographed like a product shot, and one panel of the location. Black background, no text on it. The model reads that sheet alongside your prompt and keeps everybody looking like themselves.
One tip from their own documentation that is easy to miss: the more space something takes up on your sheet, the more faithfully it comes through. Give the character who matters a big panel. A crowded sheet gives you drift.
The seven that stayed behind
Now compare the older LTX 2.3 shelf against the 2.5 shelf. Seven tools exist for 2.3 and have no 2.5 version as of this morning:
- In-Outpainting. Mask a region and fill it, or extend the frame past its original edges. The one I would miss most.
- Relight. Relighting for exterior shots, steered by a light-direction ball you composite in rather than by a text description.
- DubIt. Dubbing.
- HDR. Conversion and generation for displays that show a wider range of brightness.
- Foley-V2A. Sound effects generated from the picture.
- Instant-Shave. Beard removal, which sounds like a joke until you have matched a continuity error across a reshoot.
- Cross-Eyed. Worth naming precisely because the name misleads: this is not stereo 3D. It gives a close-up portrait permanently crossed eyes. A novelty, and the one item here nobody will miss.
That list is not a footnote. If your work depends on opening up a 4:3 archive clip to 16:9, or on generating sound for a silent render, upgrading your base model costs you the tool. Nobody announces this. The old tools are not deleted, they just do not reappear, and you only notice if you go looking.
Porting them is not a formality either. Reading the In-Outpainting documentation makes that obvious: it runs in two passes, a rough fill followed by a second pass that blends the original footage back in at the seam. Its advice on masks is the kind you only learn by shipping. Paint your mask wider than the object you are removing, and include its shadow, its reflection and anywhere it touches the ground. Their reasoning: the model understands cause and effect in a scene, so if you leave the shadow, it may put the object back purely because the shadow implies it should be there.
I do not read the gap as neglect. Eight tools in three days is fast work and the queue is probably not finished. But "probably coming" is not something you can schedule a delivery around, and no published roadmap says which of the seven are next.
Where Ripple fits, and what it will not tell you
Ripple is not from Lightricks. It comes from an individual, and it does something none of the official nine do: it takes your instruction as a picture instead of a sentence.
You supply the video and a replacement for its first frame. The fifteen example clips in the repository show object replacement, outfit swaps, face swaps, texturing a grey 3D model, and a style change applied to one actor in a two-shot while the other is left alone.
My favourite example is not a listed feature at all. Because the edited first frame happened to be sharper than the source, the whole video came back sharper. The tool is not recognising a category of edit. It is reading the difference between two pictures and carrying whatever it finds through the shot.
The strength dial sits at 1.35 by default, higher than most such tools want, and the guidance is refreshingly plain: turn it up when your edit is not carrying through, turn it down when things change that should have stayed put.
Here is the honest part. Ripple publishes nothing about how it was made. No training notes, no timings, no memory requirement, not even which version of the base model it expects. You get the file, a workflow and fifteen videos.
That matters practically, not academically. A tool with no documentation is an experiment, not something to build a delivery around, because the only way to find out where it fails is to hit that failure on a real shot.
Try it in about thirty minutes
- Install the node pack. In ComfyUI Manager, search "LTXVideo" and install ComfyUI-LTXVideo. Use the IC-LoRA loader node that comes with it. A generic LoRA loader will appear to work while ignoring your input clip entirely, which is the single most common way this goes wrong.
- Accept the licence gates. The LTX 2.5 model files sit behind a click-through on Hugging Face. A free account clears it. Be realistic about the download; this is tens of gigabytes.
- Start with Clean-Plate, not Ripple. Run it on a clip where you want one thing gone. It is the cheapest way to find out whether your machine can handle any of this, and the result is unambiguous. The object is either gone or it is not.
- Then try Ripple. Export your first frame, edit it, load it back in beside the original clip, set strength to 1.35, and leave the default prompt alone on the first run so you can see what the tool does without your wording interfering.
- Run the upgrade check before you commit. Open the Lightricks page, list the tools you actually use, and confirm each one exists for the version you plan to run. Five minutes now, or a rebuilt pipeline later.
What to watch out for
The entry cost is front-loaded. The add-ons are small. The base model they all need is not, and it sits behind an account and a licence acceptance. You will spend an evening downloading before you generate a frame.
There is a revenue cliff in the licence. A company turning over 10 million dollars a year or more needs a paid agreement with Lightricks before shipping work made with LTX 2.x or anything built on it. That is in the licence file every one of these tools points at, and it is exactly the sort of thing a legal team finds after the shot is approved.
Trust the repair tools least. Deblur, Decompression and the upscaler are demonstrated on damage they were trained to reverse, and your footage is damaged differently. A useful cautionary result surfaced this week from someone working on an entirely different restoration model: on his own test clips, a correct, carefully verified version still looked worse than a plain resize. Before you adopt any of these, put three versions side by side at full size, your source, the model's output, and a simple resize, and look. It takes twenty minutes and it settles the question.
Ripple's missing documentation has a practical edge. Without a stated base version you will not know whether to run the fast distilled model or the full one, and the settings differ enough between them that a bad first result tells you nothing at all.
The check that actually matters
What sticks with me is that this shelf carries no promise. Nine tools exist for 2.5 because someone spent three days porting them, and seven do not because that work has not happened yet, or will not. Two of the most interesting ones came from individuals, one of whom documented everything and one of whom documented nothing.
That is a healthy scene and a fragile supply chain at the same time. So the habit worth building is not picking the best tool. It is knowing which of the ones you rely on would genuinely hurt to lose, and whether anything about how they were made suggests they will still be around in six months.
Go look at your own list. If everything on it is official and ported, you are fine. If the one you cannot work without is a file from an account with no other releases and no notes, better to know that now than on a deadline.
And if you have run Ripple on real footage, I want to hear where it broke. The failure modes are the part nobody has written down, and right now the only way that gets written is if the people using it say what happened.
- Canonical sources: LTX Ripple · Lightricks on Hugging Face · ComfyUI-LTXVideo · LTX IC-LoRA docs · LTX-2.x Community License