LEJ

LTX 2.5 vs Wan 2.2 image to video: one still per prompt, ten prompts, one Spark

Rated G

On one seed, LTX 2.5 followed the prompt more often and rendered about four times faster; Wan 2.2 had the most physical motion and the biggest invention, a glass carafe that pours into the mug.

Instructional (3 min 28 s): how we ran the head-to-head, hosted by three AI hosts with AI voices (Daisy Lane, Rafael Duarte, Priya Vance). On-camera takes are LTX 2.5 image+audio renders; lip sync is loose at pauses, so the edit cuts to b-roll over the longer ones. Music: ElevenLabs Music, AI-generated.
Prompt and settings

Settings: Remotion Instructional, 1920x1080, 30 fps; host takes LTX 2.5 ia2v 1280x704 seed 42; voices ElevenLabs eleven_v4 / eleven_v3

Reference images (3)

Listen

Transcript

Hey, I'm Priya Vance, an AI host, and this is the Video Lab. This one's a head-to-head. Here's the setup. Ten prompts, the same ten we use for every video test: a talking host, a coffee mug, a lighthouse, a skateboarder, hands kneading dough, a neon sign, a cartoon fox, a robot mascot, a rainy crosswalk, and a wet dog. For each one we made a still with Z-Image Turbo. Then we gave that exact still, the same prompt and the same seed to two image-to-video models on one DGX Spark. LTX 2.5, the twenty-two billion parameter distilled model, in int8. And Wan 2.2 image-to-video, the A14B model, in fp8 with the four-step Lightning LoRA. Both at 1280 by 704. Both about five seconds. Okay, here's the real version of the result. On seed forty-two, LTX scored higher overall on four prompts, Wan on one, and five tied. Only one gap was bigger than a point. That gap is the coffee mug. The prompt says push in slowly and nothing else moves. LTX did exactly that. Wan made the best steam of the night, skipped the push-in, and then a glass carafe slid into frame and poured something into the mug. LTX also took the robot, because it tilted its head and then waved, in that order. Wan waved and barely tilted. And it took the neon sign and the cartoon fox, each by a point. Wan took the kickflip, by a point, and honestly neither is good. Both riders float. LTX hangs in the air for almost two seconds, and a swoosh-style logo appears on the hoodie. Wan floats for about a second and a half. The ties are worth a look too. The wet dog tied overall, but Wan's shake has real weight, and the coat stays wet. On LTX the coat dries out in about a second. On the crosswalk, Wan kept the camera still like the prompt asked. LTX dollied back. Speed. LTX took about eighty-one seconds per clip, which is sixteen seconds of compute for every second of video. Wan took about three hundred thirty-five seconds per clip, sixty-six seconds per second. That's about four times slower. Wan also writes sixteen frames a second, and no audio. LTX writes twenty-four, with a generated ambient track. Stay with me, because this matters. Every comparison here is one seed. A one-point gap on one seed is noise. The second-seed runs were planned, then dropped from today's queue, so treat the close calls as ties. So here's how I'd use this. For a five-second shot from a still, LTX is the default on this box. It's four times faster and it followed the prompts more often. When the shot is about motion itself, fur, water, a shake, render the Wan take and compare. And before you trust any of this for your shot, render two seeds. Every number comes from a timing file, and every score comes from looking at frames. The clips are on the page. Watch them before you believe me. I'm Priya. See you in the lab.

Short film, 'Same Still, Two Models' (43 s): the lighthouse from both models back to back, then the higher-scoring clip for seven more prompts (on a tie, the better motion or framing; the talking host and the neon sign are left out); the caption on each shot names the model. Voice-over by Priya Vance, an AI host with an AI voice (ElevenLabs). Music: ElevenLabs Music, AI-generated.
Prompt and settings

Settings: Remotion ShortFilm, 1920x1080, 30 fps; clips as rendered (Wan conformed to 24 fps)

Reference images (10)
P02 Coffee mug push-in: LTX 2.5 (left) and Wan 2.2 (right) from the same still, prompt and seed 42. Overall 4/5 vs 2/5. Wall 78.1 s vs 357.3 s (15.5 vs 70.6 s per second of video). Wan conformed from 16 to 24 fps by frame duplication; LTX ambient audio at -12 dB. Tests 022 and 047.
Prompt and settings

Prompt: A matte black ceramic coffee mug on a walnut table, thin steam rising and curling in a shaft of morning window light. The camera pushes in slowly toward the mug. Nothing else moves. Photorealistic, 85mm lens look, warm light.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (3)
P08 Robot: tilt, then wave: LTX 2.5 (left) and Wan 2.2 (right) from the same still, prompt and seed 42. Overall 4/5 vs 3/5. Wall 87.1 s vs 318.2 s (17.3 vs 62.9 s per second of video). Wan conformed from 16 to 24 fps by frame duplication; LTX ambient audio at -12 dB. Tests 028 and 053.
Prompt and settings

Prompt: A small round white robot mascot with glowing blue eyes sits on a wooden desk next to a laptop, tilts its head curiously, then raises one arm and waves at the camera. Static camera, soft studio light, clean 3D render look.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (3)
P04 Kickflip: LTX 2.5 (left) and Wan 2.2 (right) from the same still, prompt and seed 42. Overall 2/5 vs 3/5. Wall 78.1 s vs 318.3 s (15.5 vs 62.9 s per second of video). Wan conformed from 16 to 24 fps by frame duplication; LTX ambient audio at -12 dB. Tests 024 and 049. A swoosh-style logo LTX invented on the hoodie is blurred from 3.5 s.
Prompt and settings

Prompt: A skateboarder in a grey hoodie rides toward the camera on a concrete plaza and pops a kickflip, the board spinning under his feet before he lands and rolls past. Tracking camera, bright overcast daylight, photorealistic, sharp.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (3)
P10 Wet dog shake: LTX 2.5 (left) and Wan 2.2 (right) from the same still, prompt and seed 42. Overall 4/5 vs 4/5. Wall 81.1 s vs 315.3 s (16.1 vs 62.3 s per second of video). Wan conformed from 16 to 24 fps by frame duplication; LTX ambient audio at -12 dB. Tests 030 and 055.
Prompt and settings

Prompt: A wet golden retriever standing on a lake dock shakes the water off its coat, droplets flying in every direction, ears flapping, in bright afternoon sun. Static camera at the dog's eye level, photorealistic, slow motion feel.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (3)
P06 Neon sign text: LTX 2.5 (left) and Wan 2.2 (right) from the same still, prompt and seed 42. Overall 4/5 vs 3/5. Wall 84.1 s vs 354.3 s (16.7 vs 70 s per second of video). Wan conformed from 16 to 24 fps by frame duplication; LTX ambient audio at -12 dB. Tests 026 and 051.
Prompt and settings

Prompt: A pink neon sign on a dark brick wall that reads VIDEO LAB in bold capital letters flickers twice and then glows steadily, soft pink light spilling onto the bricks. Static camera, night, photorealistic.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (3)
P07 Cartoon fox wave: LTX 2.5 (left) and Wan 2.2 (right) from the same still, prompt and seed 42. Overall 5/5 vs 4/5. Wall 84.2 s vs 339.2 s (16.7 vs 67 s per second of video). Wan conformed from 16 to 24 fps by frame duplication; LTX ambient audio at -12 dB. Tests 027 and 052.
Prompt and settings

Prompt: A cheerful stylized 3D cartoon fox with big eyes stands in a sunny storybook forest clearing, looks at the camera and waves with one paw, tail swishing, leaves drifting down. Static camera, bright saturated colors, Pixar-like render.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (3)
P03 Lighthouse orbit: LTX 2.5 (left) and Wan 2.2 (right) from the same still, prompt and seed 42. Overall 4/5 vs 4/5. Wall 78.1 s vs 351.3 s (15.5 vs 69.4 s per second of video). Wan conformed from 16 to 24 fps by frame duplication; LTX ambient audio at -12 dB. Tests 023 and 048.
Prompt and settings

Prompt: Aerial drone shot orbiting slowly around a lighthouse on a rocky coastal cliff at golden hour, waves breaking white against the rocks below, long shadows, clouds drifting. Smooth steady orbit, wide lens, photorealistic.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (3)
P09 Rainy night crosswalk: LTX 2.5 (left) and Wan 2.2 (right) from the same still, prompt and seed 42. Overall 4/5 vs 4/5. Wall 81.2 s vs 315.3 s (16.1 vs 62.3 s per second of video). Wan conformed from 16 to 24 fps by frame duplication; LTX ambient audio at -12 dB. Tests 029 and 054.
Prompt and settings

Prompt: A busy city crosswalk at night in the rain, dozens of pedestrians with umbrellas crossing, car headlights and neon shop signs reflecting in the wet asphalt. Static camera at street level, cinematic, photorealistic.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (3)
P05 Hands kneading dough: LTX 2.5 (left) and Wan 2.2 (right) from the same still, prompt and seed 42. Overall 4/5 vs 4/5. Wall 81.2 s vs 330.2 s (16.1 vs 65.2 s per second of video). Wan conformed from 16 to 24 fps by frame duplication; LTX ambient audio at -12 dB. Tests 025 and 050.
Prompt and settings

Prompt: Close-up of two hands kneading bread dough on a flour-dusted wooden counter, fingers pressing and folding the dough, flour puffing up slightly. Static camera from above at a slight angle, warm kitchen light, photorealistic.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (3)
P01 Talking host: LTX 2.5 (left) and Wan 2.2 (right) from the same still, prompt and seed 42. Overall 4/5 vs 4/5. Wall 78.2 s vs 366.3 s (15.5 vs 72.4 s per second of video). Wan conformed from 16 to 24 fps by frame duplication; LTX ambient audio at -12 dB. Tests 021 and 046.
Prompt and settings

Prompt: A young woman with blonde hair in a pink crop top sits on a stool in a bright modern studio and talks to the camera with a warm, animated smile, small natural head movements and hand gestures. Chest-up framing, eye level, static camera, soft key light, shallow depth of field, photorealistic.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (3)
P02 Coffee mug push-in, LTX 2.5 alone, as rendered, 24 fps, with LTX's generated ambient audio. Overall 4/5; 78.1 s wall. Test 022.
Prompt and settings

Prompt: A matte black ceramic coffee mug on a walnut table, thin steam rising and curling in a shaft of morning window light. The camera pushes in slowly toward the mug. Nothing else moves. Photorealistic, 85mm lens look, warm light.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (2)
P02 Coffee mug push-in, Wan 2.2 alone, as rendered, 16 fps; Wan writes no audio, so this file carries a silent track. Overall 2/5; 357.3 s wall. Test 047.
Prompt and settings

Prompt: A matte black ceramic coffee mug on a walnut table, thin steam rising and curling in a shaft of morning window light. The camera pushes in slowly toward the mug. Nothing else moves. Photorealistic, 85mm lens look, warm light.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (2)
P08 Robot: tilt, then wave, LTX 2.5 alone, as rendered, 24 fps, with LTX's generated ambient audio. Overall 4/5; 87.1 s wall. Test 028.
Prompt and settings

Prompt: A small round white robot mascot with glowing blue eyes sits on a wooden desk next to a laptop, tilts its head curiously, then raises one arm and waves at the camera. Static camera, soft studio light, clean 3D render look.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (2)
P08 Robot: tilt, then wave, Wan 2.2 alone, as rendered, 16 fps; Wan writes no audio, so this file carries a silent track. Overall 3/5; 318.2 s wall. Test 053.
Prompt and settings

Prompt: A small round white robot mascot with glowing blue eyes sits on a wooden desk next to a laptop, tilts its head curiously, then raises one arm and waves at the camera. Static camera, soft studio light, clean 3D render look.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (2)
P04 Kickflip, LTX 2.5 alone, as rendered, 24 fps, with LTX's generated ambient audio; the swoosh-style logo it invented is blurred from 3.5 s. Overall 2/5; 78.1 s wall. Test 024.
Prompt and settings

Prompt: A skateboarder in a grey hoodie rides toward the camera on a concrete plaza and pops a kickflip, the board spinning under his feet before he lands and rolls past. Tracking camera, bright overcast daylight, photorealistic, sharp.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (2)
P04 Kickflip, Wan 2.2 alone, as rendered, 16 fps; Wan writes no audio, so this file carries a silent track. Overall 3/5; 318.3 s wall. Test 049.
Prompt and settings

Prompt: A skateboarder in a grey hoodie rides toward the camera on a concrete plaza and pops a kickflip, the board spinning under his feet before he lands and rolls past. Tracking camera, bright overcast daylight, photorealistic, sharp.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (2)
P10 Wet dog shake, LTX 2.5 alone, as rendered, 24 fps, with LTX's generated ambient audio. Overall 4/5; 81.1 s wall. Test 030.
Prompt and settings

Prompt: A wet golden retriever standing on a lake dock shakes the water off its coat, droplets flying in every direction, ears flapping, in bright afternoon sun. Static camera at the dog's eye level, photorealistic, slow motion feel.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (2)
P10 Wet dog shake, Wan 2.2 alone, as rendered, 16 fps; Wan writes no audio, so this file carries a silent track. Overall 4/5; 315.3 s wall. Test 055.
Prompt and settings

Prompt: A wet golden retriever standing on a lake dock shakes the water off its coat, droplets flying in every direction, ears flapping, in bright afternoon sun. Static camera at the dog's eye level, photorealistic, slow motion feel.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (2)
P06 Neon sign text, LTX 2.5 alone, as rendered, 24 fps, with LTX's generated ambient audio. Overall 4/5; 84.1 s wall. Test 026.
Prompt and settings

Prompt: A pink neon sign on a dark brick wall that reads VIDEO LAB in bold capital letters flickers twice and then glows steadily, soft pink light spilling onto the bricks. Static camera, night, photorealistic.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (2)
P06 Neon sign text, Wan 2.2 alone, as rendered, 16 fps; Wan writes no audio, so this file carries a silent track. Overall 3/5; 354.3 s wall. Test 051.
Prompt and settings

Prompt: A pink neon sign on a dark brick wall that reads VIDEO LAB in bold capital letters flickers twice and then glows steadily, soft pink light spilling onto the bricks. Static camera, night, photorealistic.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (2)
P07 Cartoon fox wave, LTX 2.5 alone, as rendered, 24 fps, with LTX's generated ambient audio. Overall 5/5; 84.2 s wall. Test 027.
Prompt and settings

Prompt: A cheerful stylized 3D cartoon fox with big eyes stands in a sunny storybook forest clearing, looks at the camera and waves with one paw, tail swishing, leaves drifting down. Static camera, bright saturated colors, Pixar-like render.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (2)
P07 Cartoon fox wave, Wan 2.2 alone, as rendered, 16 fps; Wan writes no audio, so this file carries a silent track. Overall 4/5; 339.2 s wall. Test 052.
Prompt and settings

Prompt: A cheerful stylized 3D cartoon fox with big eyes stands in a sunny storybook forest clearing, looks at the camera and waves with one paw, tail swishing, leaves drifting down. Static camera, bright saturated colors, Pixar-like render.

Settings: Image-to-video from a 1344x768 Z-Image Turbo still (seed 42), both models at 1280x704. LTX 2.5 22B distilled int8: 121 frames at 24 fps, two-stage 8+3 steps, CFG 1.0, seed 42. Wan 2.2 I2V A14B fp8 + lightx2v 4-step LoRAs: 81 frames at 16 fps, 4 steps (2 high-noise + 2 low-noise), shift 5, CFG 1.0, seed 42.

Reference images (2)

P02 Coffee mug push-in: LTX 2.5 vs Wan 2.2, same still, seed 42

P08 Robot: tilt, then wave: LTX 2.5 vs Wan 2.2, same still, seed 42

P04 Kickflip: LTX 2.5 vs Wan 2.2, same still, seed 42

P10 Wet dog shake: LTX 2.5 vs Wan 2.2, same still, seed 42

P06 Neon sign text: LTX 2.5 vs Wan 2.2, same still, seed 42

P07 Cartoon fox wave: LTX 2.5 vs Wan 2.2, same still, seed 42

Scrub LTX 2.5 and Wan 2.2 side by sideTen prompts, one Z-Image still each, seed 42. Pick a prompt; one scrubber drives both clips. LTX 2.5 22B distilled on the left, Wan 2.2 I2V A14B with the 4-step LoRA on the right.
Reference images (10)
The scorecard: sort by any scoreAll 20 clips scored 1-5 from frames on seed 42. Click a column to sort, click a row for a frame and a short note. One seed per prompt, so a one-point gap is noise.
Reference images (10)

Verdict

Run LTX 2.5 as the default image-to-video model on a DGX Spark, and render a Wan 2.2 take when the shot is about the motion itself. On seed 42 across ten prompts, LTX scored higher overall on 4, Wan on 1, and 5 tied. Only the coffee mug was a gap bigger than one point: Wan skipped the push-in and invented a carafe that pours into the mug. LTX took a median 16.1 s of compute per second of video, Wan 66.1 s, about 4.1 times as long. Wan writes 16 fps and no audio, so conform it before cutting. One seed is one draw: treat the one-point calls as ties.

Credits and settings

ModelLTX 2.5 22B distilled vs Wan 2.2 I2V A14B
MakerLightricks (LTX 2.5); Alibaba Tongyi Lab / Wan-AI (Wan 2.2); lightx2v (4-step LoRAs); Comfy-Org repacks
VersionLTX 2.5 22B distilled (int8, 2-stage) vs Wan 2.2 I2V A14B (fp8, lightx2v 4-step LoRA v1)
What it doesimage-to-video
LicenseLTX-2.x Community License Agreement (LTX 2.5); Apache-2.0 (Wan 2.2 I2V A14B, the lightx2v LoRAs, umt5-xxl and the Wan 2.1 VAE as repackaged by Comfy-Org); Apache-2.0 (Z-Image Turbo, the source stills)
HardwareDGX Spark: NVIDIA GB10, aarch64, 128 GB unified memory (about 121 GB visible). ComfyUI 0.43.0. LTX on one ComfyUI instance, Wan on a second; the other instance's models were unloaded before each batch. A 19 GB voice model and a third ComfyUI were resident during the LTX runs; their state during the Wan runs was not recorded. By the Wan runs, both lab ComfyUIs ran inside a memory-capped service group (85 GB shared) that did not exist during the LTX runs.
Speed16 s of compute per second of video. LTX 2.5: median 16.1 s of compute per second of 1280x704 video (15.5-17.3, tests 021-030, 78.1-87.1 s per 5.04 s clip). Wan 2.2 with the 4-step LoRA: median 66.1 s (62.3-72.4, tests 046-055, 315.3-366.3 s per 5.06 s clip). The families ran 11 hours apart, across a kernel and NVIDIA driver change.
Peak memoryPeak use not captured. Lowest MemAvailable sampled every 2 s: 13 GB during the LTX runs (test 028), 23 GB during the Wan runs (test 051), under different background load. 128 GB unified memory, about 121 GB visible.
Tested onNov 2, 2026
AI hosts

Daisy Lane (AI host; ElevenLabs Jessica (AI voice), AI voice)

Priya Vance (AI host; ElevenLabs designed voice (AI), AI voice)

Rafael Duarte (AI host; ElevenLabs designed voice (AI), AI voice)

Narration
Priya Vance (AI host, ElevenLabs voice) (AI voice, ElevenLabs)
Music source
ElevenLabs Music, AI-generated instrumental bed
Production
LEJ Video Lab burn run, Nov 2 2026, on one DGX Spark. Scores from frames by Claude reviewers; every number from the test folders' timing files.

What it does well

  • LTX 2.5 did the named beats in order where Wan did not: the slow push-in on the mug (022) and the tilt-then-wave robot (028).
  • LTX 2.5's cartoon fox (027) was the only 5/5 overall in the set: a wave, a wink and a swishing tail, on-model throughout.
  • LTX 2.5 was about four times faster than Wan 2.2 per second of video and writes its own ambient audio track.
  • Wan 2.2's wet dog shake (055) is the best motion in the set: head whip, flying ears, and the coat stays wet.
  • Wan 2.2 held the static camera on the rainy crosswalk (054) where LTX dollied back, and its kickflip (049) floated less with no added logo.

Where it breaks

  • Wan 2.2 invented an object: a glass carafe enters at 1.8 s and pours into the mug, against 'nothing else moves' (047).
  • Both kickflips float: about 1.8 s of hang time on LTX (024), about 1.4 s on Wan (049); LTX also added a swoosh-style logo to the hoodie (blurred in the clips on this page).
  • Neither model orbited the lighthouse; both tilted up and pulled back (023, 048).
  • Wan's neon sign pulses four times and never settles to the steady glow asked for (051); LTX dips twice then holds (026).
  • Wan writes 16 fps and no audio; LTX writes 24 fps with a generated ambient track. Cutting them together needs a conform or interpolation.
  • One seed per prompt: the planned seed-1234 repeats were dropped, so one-point gaps are within noise.

We gave LTX 2.5 and Wan 2.2 the same ten starting pictures, the same prompts and the same seed, on one DGX Spark. LTX followed the prompt more often and rendered about four times faster. Wan made the most physical motion in the set, a wet dog's shake, and also invented a coffee carafe nobody asked for. It's one seed per prompt, so treat the one-point calls as ties.

What we tested

Ten prompts from the lab's fixed suite, each built to stress one thing:

PromptWhat it stresses
P01 talking hostface, idle motion
P02 coffee mugslow push-in, steam, nothing else moving
P03 lighthousea drone orbit
P04 kickflipfast limbs
P05 hands kneading doughfingers
P06 neon signtext that must stay legible
P07 cartoon foxa stylized look held over time
P08 robottwo beats in order: tilt, then wave
P09 rainy crosswalkmany moving people
P10 wet dogfur and water

For each prompt we made one still with Z-Image Turbo, so both models start from the exact same pixels. Then we gave that still and the motion prompt to LTX 2.5 22B distilled and to Wan 2.2 I2V A14B with the 4-step LoRA. Both rendered 1280x704 clips of about five seconds.

The lab's harness saved the exact ComfyUI graph, the seed, the wall clock and free memory for every run. Each clip was then scored 1-5 from a 12-frame sheet on five criteria: prompt adherence, motion, consistency, artifacts and overall usable for publication. LTX was scored by the lab's test reviewer and Wan by this post's producer, on the same rubric, because the reviewer was paused when the Wan clips landed. Both are Claude models reading frames, not people watching motion.

The plan had a second seed on five prompts for both models. Those runs were dropped from the day's queue to make room for other jobs, so every call here is one draw.

What happened

PromptLTX overallWan overallCall
P01 talking host44Tie
P02 coffee mug42LTX
P03 lighthouse44Tie
P04 kickflip23Wan by 1
P05 hands44Tie
P06 neon sign43LTX by 1
P07 cartoon fox54LTX by 1
P08 robot43LTX by 1
P09 crosswalk44Tie
P10 wet dog44Tie

LTX scored higher on four prompts, Wan on one, and five tied. Every sub-score and the reviewer's note for each clip are in the scorecard player on this page.

The one clear gap is the coffee mug. The prompt says push in slowly and nothing else moves. LTX did that. Wan made the best steam in the set, skipped the push-in, and at 1.8 s a glass carafe slides in and pours into the mug.

LTX was better at doing what it was told. The robot tilts its head and then waves, in that order; Wan waves and barely tilts. The neon sign dips twice and then holds; Wan's pulses four times and never settles. The fox waves, winks and swishes its tail.

Wan was better at motion. The wet dog tied overall, but Wan's shake has weight: the head whips, the ears fly and the coat stays wet, where the LTX coat dries out in about a second. On the crosswalk Wan held the static camera the prompt asked for; LTX dollied back.

Both broke on the kickflip. Both riders float far too long. LTX also put a swoosh-style logo on the hoodie, which we blurred in the clips on this page.

Speed and memory

LTX 2.5 (tests 021-030)Wan 2.2 (tests 046-055)
Wall time per clip78.1-87.1 s, median 81.2 s315.3-366.3 s, median 334.7 s
Seconds of compute per second of videomedian 16.1median 66.1
Lowest free memory sampled13 GB23 GB
Output24 fps, with an ambient audio track16 fps, no audio

Wan took about 4.1 times as long per second of video. Wall time runs from submitting the graph to ComfyUI reporting done, including the VAE decode and the mp4 encode. A separate Wan model-load cost is not measured here: the first clip was 31.6 s above the median, but the second was close behind it.

The two families ran eleven hours apart, and the box changed in between. It froze twice, rebooted, moved to a new kernel and NVIDIA driver, and put both ComfyUI servers inside a shared 85 GB memory cap. The day before, on the old kernel, the same Wan setup took 336 s per clip, within 2 s of today's median, so the kernel change doesn't explain the gap. The memory minimums were taken under different background load, so they are not a like-for-like comparison, and per-process GPU memory was not captured.

What it means for you

On a Spark, make LTX 2.5 your default for a five-second shot from a still. It's about four times faster, and on this seed it followed the prompt more often.

When the shot is about motion itself, like fur, water, steam or a body shake, render a Wan take too and compare. Watch it for invented objects.

Wan writes 16 fps and LTX writes 24, so conform or interpolate Wan before you cut the two together. And render two seeds before you trust a call like the ones above.

How to do it

One model family at a time. LTX runs on one ComfyUI instance and Wan on a second, and the harness unloads the other instance before each batch. On a GB10 the CPU and GPU share one memory pool, and with both families loaded there would be little left for anything else (we did not try it). The box froze twice on the day of this test, both times with an LTX job on the GPU and a headless browser playing video in a page. That cause is likely, not proven.

Wan 2.2 I2V is two models. A high-noise and a low-noise 14B expert, 14.3 GB each in fp8, take turns during sampling. The 4-step LoRA is what makes it usable here: the reference setup is 40 steps with guidance 3.5, and we did not run that. This is Wan 2.2 at its fast 4-step setting against LTX 2.5's distilled model at its two-stage setting. LTX 2.5 also ships a full dev model, which we did not test.

Negative prompts do nothing at CFG 1. Both graphs run at CFG 1.0, where the negative branch has no effect.

Mind the frame rates. The side-by-sides and the edits conform Wan to 24 fps by duplicating frames, which is honest but a little steppy. The players show both clips as rendered.

Settings

LTX 2.5 22B distilled. ltx-2.5-22b-distilled-transformer-comfy-int8-convrot with the int8 Gemma 4 12B text encoder, the LTX video and audio VAEs and the x2 latent upscaler. Stage one at 640x352 on the 8-step distilled schedule, then stage two at 1280x704 with 3 refinement steps. euler_ancestral, CFG 1.0, the still in at strength 1.0, 121 frames at 24 fps, seed 42.

Wan 2.2 I2V A14B, 4-step. wan2.2_i2v_high_noise_14B_fp8_scaled and wan2.2_i2v_low_noise_14B_fp8_scaled, umt5_xxl_fp8_e4m3fn_scaled, wan_2.1_vae, and the lightx2v 4-step LoRAs v1 (high and low) at strength 1.0, shift 5. Four steps in total, euler, simple, CFG 1.0: steps 0-2 on the high-noise model, 2-4 on the low-noise model. 81 frames at 16 fps, seed 42.

Stills. Z-Image Turbo, 1344x768, 8 steps, seed 42. Both video graphs resize the still to 1280x704.

The linked workflow is the LTX lighthouse graph (test 023), built from ComfyUI 0.43 core nodes.

Sources: LTX-2.5 model card, LTX-2.x Community License, Wan2.2-I2V-A14B model card and license, Wan 2.2 reference config, lightx2v Wan2.2-Lightning, Comfy-Org Wan 2.2 repackage.

About the videos

The short film shows the lighthouse from both models back to back. Then it cuts the higher-scoring clip for seven more prompts; on a tie it takes the better motion or framing. The talking host and the neon sign are left out.

The instructional is hosted by three AI hosts with AI voices (ElevenLabs): Daisy Lane, Rafael Duarte and Priya Vance. Their on-camera takes are LTX 2.5 image+audio renders from each host's studio still. Lip sync is loose at pauses (the mouth closed at 4 of 23 sampled pauses), so the edit cuts to b-roll over pauses longer than half a second. Music is ElevenLabs Music, AI-generated.

ComfyUI API workflow for the LTX lighthouse clip (test 023); the Wan settings are in the Settings section ↗