LTX-2.5's matte adapter takes 3.1 minutes on a 16 GB GPU

An independent studio has published the timings that decide whether Lightricks' LTX-2.5 post-production adapters work at home. OMNI Labs ran three of them on its own RTX 5060 Ti with 16 GB of VRAM and posted the numbers on 28 October 2026. SLOP TV ran none of these tests and has not tested the adapters.

The short version: Clean Plate removed a walking subject from a moving street shot in 2.9 minutes for 97 frames, Alpha Gen pulled a matte in 3.1 minutes, and Refine Details spent 11.2 minutes turning a 720x1280 clip into a 1088x1920 one that OMNI judged barely sharper than a plain Lanczos resize. All three peaked at 15.5 GB of VRAM.

What Lightricks shipped, and what it asks for

The adapters arrived through VFX Week, a run of daily drops that closed on 2 October 2026. Refine Details and Restore shipped on 29 September, and Alpha Gen followed on 2 October. OMNI puts the wider catalog at about twenty IC-LoRAs, add-ons that take an existing clip as a reference and return a transformed version.

Alpha Gen is the one most creators can use without a shot list. It takes an ordinary RGB clip and returns a grayscale matte, white for the subject and black for everything else, with no prompt and no mask. Lightricks' documentation caps it at 1920x1088 and calls 121 frames or fewer the reliable range; counts up to 145 are possible but less dependable, and no text prompt changes which subject the model picks.

Then the hardware floor. Lightricks' own system requirements list a minimum of 32 GB of VRAM for LTX-2.5, with an A100 at 80 GB or an H100 recommended. On 16 GB, OMNI reports that the model does not fit and part of it sits in shared system memory, and its times include that penalty. Every number below is an edge case, not a benchmark of the adapter on the hardware Lightricks expects.

Where the tests agree

OMNI kept Clean Plate and Alpha Gen and dropped Refine Details. Clean Plate removed the walker, her handbag and her shadow with no ghosted outline, and the facade revealed behind her moved with the camera. Alpha Gen chose the seated host and left the desk, chair, microphone and phone black, with the gaps between her arms open and her hair soft at the edge.

Loop Forge, testing separately on an RTX 3060 with 12 GB, reached a similar place. It ran 14 official Lightricks LoRAs across 31 tests and reported that "Clean Plate, Colorization, Decompression, Deblur, Restore and Alpha Gen worked on every clip I gave them," including archive footage. Peak VRAM across its tests was 7.5 to 11.8 GiB, and each clip took 3 to 36 minutes.

Refine Details had no such defender. OMNI compared the same frame at 3x zoom against Lanczos, a browser upscaler and FlashVSR+, and ranked the LTX result last of the four on sharpness. FlashVSR+, which OMNI times at about a minute per second of footage, kept lashes and earring detail that Refine Details smeared.

Where they split

The timings diverge because the clips were different sizes. OMNI's were 97 frames, roughly four seconds at 24 fps, at 704x1280 and 720x1280, run in int8 at 8 steps, cfg 1 and adapter strength 1.0. The 97 frames were OMNI's own ceiling, because longer LTX runs had hung in VAE decode on that card.

Loop Forge's were about five seconds, mostly 121 frames, on a 12 GB card. It still peaked lower, at 11.8 GiB, while taking up to 36 minutes for a single clip. Neither lab published a per-adapter table for every test, so the ranges do not line up cleanly.

At the other end, a compositor on the Logik forum, the Autodesk Flame community, ran Alpha Gen at 3200x1900 for 49 frames in bf16 and measured about 90 GB of VRAM and about 2.5 minutes. CreativeAiNews dates the post to 3 October. That resolution is above the 1920x1088 the model card allows, so it is a stress test rather than a like-for-like figure. A second poster in the thread ran seven hand-movement clips, mostly 70 frames at 1920x1080, many in 64 seconds, and one at the full 145 frames in UHD took 39 minutes.

The Refine verdict needs a caveat. OMNI is candid that its test sat outside the adapter's training: Refine Details was built to take 1080p to 4K, and a 1.5x step puts each tile outside the scale it learned. The lab also ran int8 where the model card assumes bf16. ComfyUI Wiki notes the card recommends refining to 4K first and resizing to 8K, because a direct 8K pass can invent texture. That explains the weak result without overturning it, and OMNI estimates well over half an hour for the true 1080p-to-4K run, which it did not attempt.

The Logik thread also splits on matte quality against other tools. One poster preferred SAMMIE for consistency, and gave LTX the edge only on hair. Another said Alpha Gen beat a key on a green-screen job with blowing hair. Neither published a controlled comparison.

What it costs

There is no per-run price to compare, because the adapters are free below a revenue line. Alpha Gen and the rest fall under the LTX-2 community license, royalty-free below $10 million in annual revenue and requiring a paid commercial license above it. OMNI puts its own cost at $0 per clip. Loop Forge paid nothing either.

The cost is the card and the clock. At OMNI's settings, a four-second vertical clip takes about three minutes of a 16 GB GPU for a clean plate or a matte, and 11.2 minutes for an upscale that did not beat a free resize. At Loop Forge's settings, one clip can take 36 minutes. The Logik compositor's 49-frame stress test needed about 90 GB, which is H100 territory.

Who each test suits, and what is still open

OMNI's verdict is narrow and practical. Clean Plate suits a subject who leaves background around them; the model card puts subjects that fill the middle of the frame out of scope, and OMNI skipped its own seated close-up for that reason. Alpha Gen suits a shot with one clear subject; with two people, it cannot be told which to keep. Both now sit in OMNI's pipeline for vertical social delivery at about three minutes a clip.

Loop Forge's results point at a 12 GB card and patience. Six of its 31 tests failed, which it presents as the places a LoRA reached its limit, and chaining two LoRAs worked for Clean Plate with Day-to-Night but failed for Ingredients with Slow Motion on all three seeds it tried.

The open questions are the ones nobody has run. No published test covers SDR to HDR or the native 4K and 8K path on a consumer card; OMNI skipped both. The 121-frame reliable ceiling on Alpha Gen is stated in Lightricks' documentation and was not tested past by either lab.

The adapters are on Hugging Face under the LTX-2 community license. Accept the terms on each repository separately, then load the matching ComfyUI-LTXVideo workflow with adapter strength 1.0.