LTX 2.5 Now Has Nine Editing Adapters. Seven More Never Made the Jump.

What LTX 2.5 can now do to footage you already shot, which tools you lose by upgrading from 2.3, and how to check before you move.

Take a five-second clip. Open frame zero in whatever image editor you already own. Set the building on fire, or put a different jacket on the actor, or paint the untextured grey clay model. Save it, feed it back in beside the untouched video, and the change carries through the rest of the shot. Through camera cuts. With the original motion, timing and framing intact.

That is not a new model. It is a 654 MB file called LTX Ripple that went up on Sunday afternoon, and it is the second adapter a private individual has published for LTX 2.5 in three days. It landed at the end of a week in which Lightricks moved eight of its own editing adapters onto the 2.5 base.

Something worth noticing about that: no lab I track shipped a video model in that window. The week's product was adapters, and the input they all take is footage that already exists.

Why a shelf of small files matters more than a new checkpoint

For most of this year the thing worth waiting for in open video was a checkpoint, and the question a creator asked was whether the new one beat the old one. That question has a ceiling. A better generator still starts from nothing, which means every shot is a negotiation with a prompt about a scene you have never seen.

An in-context LoRA inverts that. You hand it a video and it hands back the same video, altered. The composition is yours. The camera move is yours. The performance is yours. The model's job is narrow, which is why a 654 MB adapter can do it and a 22B checkpoint on its own cannot.

That narrowness is the whole design. An IC-LoRA does not learn a look. During training it sees pairs: a control input and the result it should produce from that input. At inference you supply a new control and it produces the matching result. Lightricks documents this in its IC-LoRA usage guide, and the pattern is general enough that the control can be almost anything. A binary mask. A grey clay render. A black-and-white frame. A reference sheet. An edited first frame.

So the interesting question stopped being "how good is LTX 2.5" and became "what controls exist for it." As of this morning, nine.

What is on the 2.5 shelf right now

Between September 8 and September 10, Lightricks published eight adapters for LTX 2.5, joining the Pixel Spatial Upscaler that arrived on August 11. Here is the shelf, with what each one is actually for:

Adapter	What it takes in	What you get
Clean-Plate	your shot	the shot with an object removed
Day-To-Night	your shot	the same shot relit as night
Colorization	black and white footage	colour
Deblur	soft footage	sharpened footage
Decompression	compressed footage	artefacts removed
Pixel Spatial Upscaler	small footage	bigger footage
Water-Simulation	a shot	water added
Ingredients	a reference sheet	a new clip using those characters and props
Cinemagraph	a still image	selective motion

Downloads tell you which ones people actually reach for, and it is not the restoration tools. Ingredients leads the new batch at 490 as of this morning, Clean-Plate at 124, Water-Simulation at 113. Ingredients is the outlier because it solves the problem that makes generative video useless for narrative work: the same character looking like a different person in every shot. You build one composite sheet with a face close-up, a body turnaround, each prop rendered product-style, and one clean location panel, all on black with no text. The model reads that in context and keeps everything consistent.

The Ingredients card is also the most useful document on the shelf, because it tells you what the adapter cost to make: rank 128, alpha 128, bf16, AdamW-8bit, 12,000 steps, eight GPUs running DDP, trained at a single bucket of 768x448, 121 frames, 24 fps. Hold that number. It comes back later.

The seven that stayed behind

Now count the 2.3 shelf against the 2.5 shelf on the same account page.

LTX 2.3 has adapters that 2.5 does not. Seven of them, as of this morning:

• In-Outpainting. Mask a region and fill it, or extend the canvas past its original edges.
• Relight. Relighting for exterior shots, steered by a composited light-direction ball rather than by a text description.
• DubIt. Dubbing.
• HDR. Both 8-bit to 16-bit conversion and text or image driven HDR generation.
• Foley-V2A. Generate sound effects from video.
• 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 tag misleads: this is not stereo pair generation. The card describes turning a close-up portrait's straight eyes into permanent convergent strabismus. A novelty effect, and the one item on this list nobody is going to miss.

That list is not a footnote. If your pipeline depends on outpainting a 4:3 archive clip to 16:9, or on generating Foley for a silent render, upgrading your base model costs you the tool. Nobody announces that. The adapters do not get deleted, they just do not reappear, and the absence is only visible if you go looking for it.

The In-Outpainting card is the one I would miss most, and reading it shows why porting these is not trivial. It is not a single-pass adapter. It runs two stages: a coarse generation over the masked region, then a second pass with refined boundary handling that blends the original content back via a Laplacian blend. The card's mask guidance is the kind of thing you only learn by shipping. Dilate the mask past the visible edge of what you are removing, and include shadows, reflections, cast light and contact areas, because, in their words, the model understands scene causality and may re-inpaint the object from that cue alone if you leave the shadow in the frame.

That is a two-stage pipeline with workflow JSONs, blending logic and hard-won documentation attached. Reproducing it against a new base is real work, not a re-run.

I do not read the gap as neglect. The 2.5 port started on September 8 and shipped eight adapters in three days, which is a fast pace for that kind of work. The most likely reading is that the queue is not finished. But "probably coming" is not a plan you can schedule a delivery around, and there is no published roadmap that says which of the seven are queued or when.

Where Ripple fits, and what it does not tell you

Ripple is not from Lightricks. It is from an individual, credited on the card as WepeNerd, and it does something none of the official nine do: it takes your instruction as a picture instead of a sentence.

The card calls the design First Frame All Frames. You supply the video and a replacement for frame zero. The fifteen example clips in the repo show object replacement, outfit swaps, face swaps, retexturing a grey clay archviz model, and style transfer applied to one actor in a two-shot while the other is left alone. One example is the most interesting thing in the repo and is not a listed feature at all: because the edited first frame happened to be sharper than the source, the entire video came back sharpened. The adapter is not reading a category of edit. It is reading the difference between two images and propagating whatever it finds.

The default prompt ships in the workflow and starts "Use the reference video for motion, timing, camera movement, composition, and unchanged scene content." Recommended strength is 1.35, which is higher than most adapters want, and the card is blunt about how to use the dial: raise it when the edit is not carrying through, lower it when things change that should have stayed put.

Here is the honest part. Ripple publishes no training documentation at all. No rank. No dataset. No step count. No timings. No VRAM figure. Nothing about which LTX 2.5 checkpoint it was trained against, which matters because the distilled and non-distilled transformers behave differently under the same adapter. You get weights, a workflow JSON and fifteen videos.

Compare that to what Lightricks published for Ingredients, or to what AInVFX published for its fluid adapter two days earlier, which includes the full three-stage training chain, the dataset provenance down to individual Pexels video IDs, and a wall-clock figure of 7 hours 47 minutes on one rented GPU.

An adapter with no documentation is an experiment, not a dependency. That is not a knock on Ripple, whose example reel covers more ground than any single adapter on the official shelf. It is a scheduling fact. You cannot predict where an undocumented adapter fails except by running it, so you find out on the shot, not before it.

Put this into practice

The lowest-friction path is about thirty minutes, most of it downloading.

1. Install the node pack. In ComfyUI Manager, search "LTXVideo" and install ComfyUI-LTXVideo (4.1k stars). The LTX IC-LoRA Loader Model Only node comes from it. A generic LoRA loader will not work, because it ignores the reference path and silently drops the conditioning, which is the single most common way this goes wrong.
2. Accept the gates. The LTX 2.5 transformer, both VAEs and the Gemma text encoders are gated on Hugging Face. A free account and a click-through gets you through. Budget the download time honestly; this is tens of gigabytes.
3. Start with Clean-Plate, not Ripple. Download the Clean-Plate adapter into ComfyUI/models/loras/ and run it on a clip where you want one object gone. It is the cheapest way to find out whether your machine can run any of this, and the result is unambiguous. Either the object is gone or it is not.
4. Then try Ripple. Put LTX25_Ripple_v11.safetensors in the same folder, load the shipped Workflow_example/LTX_Ripple_FFAF_Edit_v11.json, export frame zero of a clip, edit it, load it back in. Set strength to 1.35. Leave the default prompt alone on the first run so you learn what the adapter does without your prompt interfering.
5. Run the upgrade check before you commit. This is the step that saves you a week. Open the Lightricks account page, list the adapters you actually use, and confirm each one exists for the base version you are planning to run. Do it now, not after you have rebuilt a pipeline around 2.5.

If step five turns up something on the missing list, you have a real decision rather than an upgrade. Running both base versions side by side is possible and it is not cheap: two 22B transformers, two sets of VAEs, and disk to match.

Honest limitations

The gating is worse than it looks on paper. The adapters are small and mostly open, but every one of them is useless without the gated base, so the practical entry cost is an account, a licence acceptance and a large download before you have generated a single frame.

The licence has a revenue cliff. The LTX-2.x Community License applies to derivatives, and a company with 10M USD or more in annual revenue needs a paid agreement with Lightricks before shipping work made with LTX-2.x or anything derived from it. That is not obscure, it is in the file every one of these adapters points at, and it is the kind of thing a studio's legal team finds after the shot is approved.

Restoration adapters are the ones I trust least. Deblur, Decompression and the upscaler are demonstrated against degradation the models were trained to reverse, and your footage is degraded differently. A conversion author working on a completely different restoration model published a useful cautionary result this week: on his own test clips, a numerically verified correct implementation still scored below plain bicubic resize on fixed quality metrics. Build the three-way comparison before you adopt any of these, source against model against bicubic, and look at all three at 100 percent.

And the specific hole in Ripple's documentation has a practical shape. Without a stated base checkpoint, you do not know whether to run the distilled transformer or the full one, and the difference in step count and sampler settings between those two is large enough to make a bad first result mean nothing at all.

The check that actually matters

The thing I keep turning over is that this shelf has no maintainer commitment attached to it. Nine adapters exist for 2.5 because someone at Lightricks spent three days porting them, and seven do not exist 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 ecosystem and a fragile supply chain at the same time, and it means the useful habit is not picking the best adapter. It is knowing which of the ones you depend on would be a genuine problem to lose, and whether anything about how they were built tells you they will still be there in six months.

Go look at your own list. If every tool on it is official, documented and ported, you are fine. If the one you cannot work without is a 654 MB file from an account with no other repos and no training notes, that is worth knowing before the deadline, not during it.

If you have run Ripple on real footage, I want to know where it broke. The failure modes are the part nobody has written down yet, and right now the only way that document gets written is if the people running it say what happened.

---

Medium metadata

• Title: LTX 2.5 Now Has Nine Editing Adapters. Seven More Never Made the Jump.
• Subtitle: What LTX 2.5 can now do to footage you already shot, which tools you lose by upgrading from 2.3, and how to check before you move.
• Tags: AI Video, ComfyUI, Open Source, Video Editing, Machine Learning
• Suggested read time: 9 minutes
• Canonical sources: LTX Ripple · Lightricks on Hugging Face · ComfyUI-LTXVideo · LTX IC-LoRA docs · LTX-2.x Community License