LTX vs Wan for image-to-video: why we parked moving scenes

Test one: Wan 2.2 on a whole story

Wan 2.2 TI2V-5B is an open image-to-video model from Alibaba under the Apache 2.0 licence, small enough to run on one mid-range data-center GPU. We wired it into the cloud renderer on an NVIDIA L4 at 480x832, 81 frames and 12 steps per clip. It runs only on NVIDIA cards.

Rather than move every scene, we moved the two that matter most: the opening, which decides whether people keep watching, and the turn, the second-to-last scene where the story changes. Everything else stayed a still with a camera move. The test story was a six-scene Urdu story about a goat called Billu.

Each moving clip took about 2.5 minutes on the L4. Measured on 9 October 2026.

Watch both versions

Here are the two versions of the same story, rendered from the same pictures. The moving version is the one on the right; watch the opening and the second-to-last scene.

Our verdict after watching them: not worth it, and not that much better. The motion is real, but at this size the movement is small and soft, and a good camera move on a sharp still already gives most of the feeling of motion. Paying 2.4 times as much and waiting more than twice as long for that was a bad trade, so we removed the feature the next day.

Test two: FastWan on a Mac and LTX-2.3 on a free demo

A week later we looked again, this time for short animated comedy scenes, and compared two models from the same starting picture of a man chasing a rooster.

FastWan 2.2 TI2V-5B is a distilled, faster version of Wan that has an MLX build for Apple silicon. On our M1 Pro a 2 second clip took 7 minutes, most of it spent decoding the frames at the end. It only gave usable output with 3 steps, a guidance scale of 1 and the Euler scheduler; the default settings did not.

LTX-2.3 from Lightricks is free to use for companies under $10 million in yearly revenue. We ran it on the free demo Lightricks hosts on Hugging Face, where each shot took about 40 seconds. Its clips were sharper than FastWan's. Its weakness was two characters close together: they can melt into one shape for a few frames. The free demo also allowed only about four shots a day, so it is fine for samples and useless for production.

An earlier warning: LTX for talking animals

This was not our first time with LTX. On 1 October 2026 we tried LTX-2.5 audio-to-video for talking animals on the Mac, using a 4-bit MLX build. The mouths and heads did move, but the result looked bad to us, and 3 seconds at 384x640 took 19 minutes.

One practical warning from that day: our first attempt at 512x896 ran with only 5 GB of free disk space and crashed the whole Mac, which restarted on its own. These models swap memory to disk at large sizes. Free up plenty of disk space before you start a big video model run on a laptop.

LTX vs Wan, side by side

Why we parked it

• Cost: moving just two scenes more than doubled the price of a video. Moving every scene would multiply it.
• Wait: a Short that took 3.5 minutes took 8.6. People notice that, and on a laptop it would be far longer.
• Look: at the sizes we can afford, the motion was not clearly better than a sharp still with a good camera move.
• Weight: the Wan model added about 22 GB to the cloud image, which slows every cold start.

What we do instead, and what would change our minds

Every TallyReel story uses stills drawn for the scene, with camera moves that give a picture a feeling of motion without a video model. Stick-figure skits get their motion from drawn poses, which costs almost nothing.

If moving moments come back, it will be as an optional extra for a few key moments rather than every scene, with the cost shown before you generate, and only after a model is clearly better than a still. A much cheaper or faster model, or one that keeps two characters apart reliably, would make us run this test again, and we will post the results here.

Models and links

• Wan 2.2 on GitHub
• LTX-2 on GitHub (Lightricks)
• FastVideo, home of FastWan (Hao AI Lab)