LEJ

Prompting LTX 2.5: camera moves, negatives, and what the model actually hears

Rated G

Fifteen LTX 2.5 clips, two scenes, two seeds, and only the words change. One camera sentence turned a parked shot into an orbit, swapping the negative prompt changed zero pixels, and over-long prompts wrecked the look. One sentence about sound changed the soundtrack and, in the kitchen, the picture.

The instructional, hosted by Daisy Lane and Theo Lindqvist, AI hosts with AI voices (ElevenLabs). Lip-sync is loose at pauses, so the edit cuts away over the longer ones. 235 s.
Prompt and settings

Settings: Hosts: LTX 2.5 image+audio from each host's on-set still, eleven_v4 voices at cast settings. Panels and grids: Remotion. Clips: the tests on this page.

Reference images (2)

Listen

Transcript

Hi, it's Theo. This one is about prompting LTX two point five, the open video model from Lightricks, and what it actually does with your words. I'm an AI host, and this is an AI voice, reading a write-up of tests the Video Lab ran on a DGX Spark on November second. Here's the setup. One model, one graph, one seed, and only the words change. We took two scenes, a lighthouse on a cliff and a black coffee mug, and wrote each one five ways. A plain sentence. The plain sentence plus one camera sentence. The camera version plus a list of negatives. An over-long prompt, the kind that gets pasted around. And the camera version plus one sentence about sound. Fifteen clips in total, five seconds each, with the two comparisons that matter most run again at a second seed. Picture it like this. Your prompt never reaches the video model as words. A twelve billion parameter language model, a version of Gemma four that Lightricks fine-tuned, reads it first, and the twenty two billion parameter video model only gets that reading. It uses it twice, in one pass: once for the picture, and once for the soundtrack. First, the camera. The plain lighthouse prompt said nothing about the camera, and the camera barely moved, at both seeds. Then we added one sentence, in the wording from Lightricks' camera guide: the camera circles around the lighthouse slowly, keeping the tower centered, and ends with the sun glowing behind it. Both seeds orbited. At the second seed, the sun traveled from the edge of the frame to directly behind the tower on the very last frame. That ending is the part worth writing: say what the frame looks like when the move is done. For the mug, we tried a film term that isn't on Lightricks' list: slow dolly in. It worked at both seeds. One pushed all the way to the rim. The other stopped about halfway. Now the negatives. Here's the trick. This distilled model runs at a guidance scale of one, and at one, ComfyUI skips the negative pass completely. So we tested it. We took the camera prompt, word for word, and swapped the negative for twelve new terms, including static camera and music. All one hundred and twenty one frames came out identical to the version with the old negative. Same pixels. The negative box did nothing. The obvious workaround is to write the no list into the prompt itself. We did that for the mug: no people, no hands, no text, no logo, no camera shake. None of those things appeared, which is good. But the whole composition changed anyway, a different chair, a different window, from the same seed. Every word you add reshuffles the frame. Then the over-long prompts. Two hundred and twenty eight words for the mug, two hundred and forty nine for the lighthouse, full of eight K, masterpiece, trending on artstation, five or six camera moves and three kinds of light. Both scenes stopped looking real. They turned into a saturated, video game style render. The camera rushed through several moves at once, the coffee changed in the middle of the shot, and both came out with loud music, probably from the one mood phrase buried in each prompt. The lighthouse one was the loudest of all twenty five clips we measured for this. Okay. Sound. Every clip LTX makes has a soundtrack, whether you ask for one or not. When the prompt says nothing about sound, the model decides. In the lab's suite of ten prompts, none of which mention audio, the talking host got a voice saying words that aren't words, and a speech detector still heard English. The cartoon fox got a beat with notes on top. And the loudness ranged across thirty one decibels, from nearly silent to peaking at full scale. So we added one sentence about sound. For the lighthouse: the deep roar of waves on the rocks, a steady wind, gulls in the distance, no music, no voices. Without that sentence, the soundtrack was a bed of held notes, more music than sea. With it, the held notes went away, and surf and wind took over, at both seeds. We couldn't find the gulls in the spectrograms. For the mug, we asked for quiet room tone, a fridge humming, a clock ticking, and a spoon clinking once against a saucer, off screen. The clip dropped almost to silence, with exactly one clink, at about three point four seconds, in both seeds. One clink, as asked. We couldn't find any ticking. And then the part nobody asked for. In both seeds, a refrigerator, a wall clock and a saucer appeared in the picture. None of them were in the picture part of the prompt. They came from the sound sentence. The model reads the whole prompt for both outputs, so if you name a sound source, expect to see it. One more practical note: the described soundtracks came out much quieter. One of them was twenty seven loudness units quieter than the same shot without a sound sentence. Normalize every clip before you cut. So here's the prompt we'd write now. Three sentences. The shot, in plain words: subject, place, light. One camera move, and what the frame looks like when it's done. And the sound, naming each source, ending with no music, no voices. Leave the negative box alone, skip the quality tags, and stop at one camera move. On speed: every clip took between seventy five and eighty one seconds for five seconds of video, about fifteen and a half seconds of compute per second, and the two hundred and forty nine word prompt was no slower than the seventeen word one. All fifteen clips, their exact prompts, the scores and the key spectrograms are on the Video Lab page. Thanks for listening.

The short film: one lighthouse, five prompts. Same model and seeds; each step changes one thing in the prompt and plays the model's own soundtrack (no music bed). 45 s. AI-generated picture and sound.
Prompt and settings

Settings: Remotion ShortFilm over tests 190, 095, 191, 096, 192, 193, 194, 097 (LTX 2.5 22B distilled, 1280x704). No music bed.

Reference images (5)
Lighthouse, plain prompt (17 words), seed 42 (test 190). Nothing about the camera: it barely moves.
Prompt and settings

Prompt: A white lighthouse on a rocky coastal cliff at golden hour, waves breaking against the rocks below.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 81.1 s (16.1 s per second of video), timing.json of test 190. Generated audio kept as the model made it.

Reference images (2)
Lighthouse, plain prompt, seed 1234 (test 095). Camera essentially locked.
Prompt and settings

Prompt: A white lighthouse on a rocky coastal cliff at golden hour, waves breaking against the rocks below.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 78.1 s (15.5 s per second of video), timing.json of test 095. Generated audio kept as the model made it.

Reference images (2)
Lighthouse + one camera sentence, seed 42 (test 191). A partial orbit: the tower stays centred while the coast swings behind it.
Prompt and settings

Prompt: A white lighthouse on a rocky coastal cliff at golden hour, waves breaking against the rocks below. The camera circles around the lighthouse slowly from left to right, keeping the tower centered, and ends with the sun glowing behind it.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 75.1 s (14.9 s per second of video), timing.json of test 191. Generated audio kept as the model made it.

Reference images (2)
Lighthouse + one camera sentence, seed 1234 (test 096). A full, clean orbit that ends with the sun directly behind the tower, as written.
Prompt and settings

Prompt: A white lighthouse on a rocky coastal cliff at golden hour, waves breaking against the rocks below. The camera circles around the lighthouse slowly from left to right, keeping the tower centered, and ends with the sun glowing behind it.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 75.1 s (14.9 s per second of video), timing.json of test 096. Generated audio kept as the model made it.

Reference images (2)
Test 191 with its negative swapped for a different 12-term list (test 192). All 121 frames are bit-identical to 191: at CFG 1.0 the negative is never read.
Prompt and settings

Prompt: A white lighthouse on a rocky coastal cliff at golden hour, waves breaking against the rocks below. The camera circles around the lighthouse slowly from left to right, keeping the tower centered, and ends with the sun glowing behind it. Negative prompt (replaces the lab default used by every other test): static camera, still frame, no camera movement, frozen, low quality, blurry, text, watermark, letterbox bars, music, voices, people

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 78.1 s (15.5 s per second of video), timing.json of test 192. Generated audio kept as the model made it.

Reference images (2)
Lighthouse, over-long prompt (249 words), seed 42 (test 193). The look flips to a saturated CG render, the camera rushes through several moves, and the soundtrack is loud music.
Prompt and settings

Prompt: An absolutely stunning, breathtaking, award-winning cinematic aerial masterpiece of a tall white and red striped historic lighthouse standing proudly on top of a dramatic, rugged, rocky coastal cliff covered in green grass, wildflowers and small stones, at golden hour with a warm orange sunset, while at the same time cold blue moonlight and the bright rotating beam of the lighthouse lamp sweep across the scene. Huge powerful turquoise waves crash and explode into white foam against the jagged black rocks far below, sea spray drifting into the air, seagulls flying in circles around the tower, a small red fishing boat sailing past on the horizon, a lighthouse keeper in a yellow raincoat walking along a winding path toward the door, and fluffy clouds racing across a sky full of pink, purple and gold. The drone camera starts with a fast handheld movement low over the water, then pushes in toward the cliff, then tilts upward to reveal the top of the lighthouse, then pulls back high into the sky and orbits around the whole headland twice, then pans across the coastline to the left. Ultra detailed, 8K, HDR, hyperrealistic, sharp focus, masterpiece, best quality, trending on artstation, volumetric god rays, lens flares, cinematic color grading, epic orchestral mood, the most beautiful lighthouse video ever made, perfectly smooth, no blur, no artifacts, no distortion, no text, no watermark, extremely high quality, professional cinematography, National Geographic style, IMAX, film grain, shallow depth of field and deep focus at the same time.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 78.1 s (15.5 s per second of video), timing.json of test 193. Generated audio kept as the model made it.

Reference images (2)
Lighthouse + camera + one sound sentence, seed 42 (test 194). Surf and wind replace the held-note bed of 191; 5.9 LU quieter.
Prompt and settings

Prompt: A white lighthouse on a rocky coastal cliff at golden hour, waves breaking against the rocks below. The camera circles around the lighthouse slowly from left to right, keeping the tower centered, and ends with the sun glowing behind it. Audio: the deep roar of waves crashing on the rocks below, a steady wind gusting across the cliff, and gulls calling in the distance. No music, no voices.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 81.1 s (16.1 s per second of video), timing.json of test 194. Generated audio kept as the model made it.

Reference images (2)
Lighthouse + camera + sound sentence, seed 1234 (test 097). Clean orbit, surf and wind, no music. The prompt we'd use.
Prompt and settings

Prompt: A white lighthouse on a rocky coastal cliff at golden hour, waves breaking against the rocks below. The camera circles around the lighthouse slowly from left to right, keeping the tower centered, and ends with the sun glowing behind it. Audio: the deep roar of waves crashing on the rocks below, a steady wind gusting across the cliff, and gulls calling in the distance. No music, no voices.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 75.1 s (14.9 s per second of video), timing.json of test 097. Generated audio kept as the model made it.

Reference images (2)
Mug, plain prompt (16 words), seed 42 (test 090). Locked camera, steam, soft window light.
Prompt and settings

Prompt: A matte black coffee mug on a walnut table, thin steam rising in morning window light.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 75.1 s (14.9 s per second of video), timing.json of test 090. Generated audio kept as the model made it.

Reference images (2)
Mug + "slow dolly in", seed 42 (test 091). A full push to the rim: a film term that isn't on Lightricks' list, and it worked.
Prompt and settings

Prompt: A matte black coffee mug on a walnut table, thin steam rising in morning window light. Slow dolly in: the camera moves slowly toward the mug until its rim and the steam fill the frame.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 78.1 s (15.5 s per second of video), timing.json of test 091. Generated audio kept as the model made it.

Reference images (2)
Mug + "slow dolly in", seed 1234 (test 098). The dolly happens again, but stops at about half the frame.
Prompt and settings

Prompt: A matte black coffee mug on a walnut table, thin steam rising in morning window light. Slow dolly in: the camera moves slowly toward the mug until its rim and the steam fill the frame.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 75.1 s (14.9 s per second of video), timing.json of test 098. Generated audio kept as the model made it.

Reference images (2)
Mug + "No people, no hands, no text, no logo on the mug, no camera shake." in the prompt (test 092). None of them appear; the composition changes anyway.
Prompt and settings

Prompt: A matte black coffee mug on a walnut table, thin steam rising in morning window light. Slow dolly in: the camera moves slowly toward the mug until its rim and the steam fill the frame. No people, no hands, no text, no logo on the mug, no camera shake.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 78.1 s (15.5 s per second of video), timing.json of test 092. Generated audio kept as the model made it.

Reference images (2)
Mug, over-long prompt (228 words), seed 42 (test 093). Every listed prop crammed in, a CG look, the coffee changes mid-shot, and lofi music.
Prompt and settings

Prompt: A gorgeous, ultra realistic, award-winning, extremely detailed commercial product video of a premium matte black ceramic coffee mug with a perfect glossy rim and a subtle embossed handle, sitting on a rich dark walnut wooden table with visible grain, knots and tiny scratches, next to a silver spoon, a croissant on a white plate, a small potted succulent, an open notebook and a pair of reading glasses, in a cozy Scandinavian kitchen. Warm golden morning sunlight streams through the window on the left while cool blue light from a neon sign and soft candlelight from the right also fill the room. Thick white steam swirls, spirals and dances upward in beautiful curls, coffee ripples gently in the cup, dust particles float in the sunbeams, a cat walks past in the background and a hand reaches in to lift the mug and then puts it back down. The camera does a slow dolly in, then a handheld movement, then circles around the mug, then tilts upward to the window, then pulls back to a wide establishing shot of the whole kitchen. 8K, HDR, hyperrealistic, masterpiece, best quality, sharp focus, bokeh, 85mm and 24mm lens, cinematic color grading, film grain, trending on artstation, the perfect coffee commercial, cozy lofi music mood, no blur, no artifacts, no distortion, no text, no watermark, highest quality possible, professional food photography, magazine cover style.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 78.1 s (15.5 s per second of video), timing.json of test 093. Generated audio kept as the model made it.

Reference images (2)
Mug + camera + one sound sentence, seed 42 (test 094). Near-silent room tone with one clink at 3.4 s, and a fridge, a clock and a saucer appear in the picture.
Prompt and settings

Prompt: A matte black coffee mug on a walnut table, thin steam rising in morning window light. Slow dolly in: the camera moves slowly toward the mug until its rim and the steam fill the frame. Audio: quiet kitchen room tone, a refrigerator humming softly, a clock ticking on the wall, and a spoon clinking once against a saucer off screen. No music, no voices.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 78.1 s (15.5 s per second of video), timing.json of test 094. Generated audio kept as the model made it.

Reference images (2)
Mug + camera + sound sentence, seed 1234 (test 099). Same result as 094: one clink at 3.4 s, and the fridge, clock and saucer show up again.
Prompt and settings

Prompt: A matte black coffee mug on a walnut table, thin steam rising in morning window light. Slow dolly in: the camera moves slowly toward the mug until its rim and the steam fill the frame. Audio: quiet kitchen room tone, a refrigerator humming softly, a clock ticking on the wall, and a spoon clinking once against a saucer off screen. No music, no voices.

Settings: LTX 2.5 22B distilled int8, ComfyUI 0.43.0, text-to-video 1280x704, 121 frames at 24 fps, two stages (8 + 3 steps), euler_ancestral, video_cfg 1.0 / audio_cfg 1.0. Wall 75.1 s (14.9 s per second of video), timing.json of test 099. Generated audio kept as the model made it.

Reference images (2)

Lighthouse, seed 42: plain, + camera, over-long, + sound

Plain vs one camera sentence, two seeds

The negative box: default vs 12-term negative (identical frames)

Mug, seed 42: plain, dolly in, "no" phrases, over-long

Lighthouse: sound described or not, two seeds

Mug: sound described or not, two seeds

Same seed, different words: LTX 2.5 prompt variants side by sidePick a comparison and scrub: every panel shows the same moment of its clip, from each clip's 12-frame contact sheet. The sound tab shows each clip's spectrogram; the clips themselves, with sound, are in the compare viewer above.
Reference images (3)

Verdict

Write LTX 2.5 prompts as three plain sentences: the shot, one camera move that says where it ends, and the sound with "No music, no voices." In our tests that gave an orbit at both seeds (one landed its end state exactly) and surf instead of a music bed, while the plain prompt left the camera parked. Don't bother with the negative box: at CFG 1.0 ComfyUI never reads it, and test 192, with its negative swapped for a different 12-term list, came out frame-for-frame identical to test 191. Skip the 200-word quality-tag prompts, which flipped both scenes to a game-engine look and added loud music. And name sound sources carefully, because the model puts them in the picture too.

Credits and settings

WorkflowPrompting LTX 2.5 22B distilled for text-to-video (picture and sound)
MakerLightricks (model); LEJ Video Lab (tests)
VersionLTX 2.5 22B distilled (Comfy int8 convrot) + Gemma 4 12B text encoder (int8), ComfyUI 0.43.0, 2 stages (8+3 steps), euler_ancestral, CFG 1.0
What it doestext-to-video prompting
LicenseLTX-2.x Community License Agreement (Aug 11, 2026)
HardwareDGX Spark: NVIDIA GB10, ARM (aarch64), 128 GB unified memory (121.7 GiB MemTotal). ComfyUI 0.43.0 (commit 426f5ee), PyTorch 2.11.0+cu130, int8 transformer and text encoder, bf16 VAEs. A voice model of about 19 GiB and a second ComfyUI were resident during the tests.
Speed16 s of compute per second of video. Median of 15 text-to-video clips at 1280x704, 121 frames (5.04 s), two stages, weights resident; range 14.9 to 16.1 s per second of video, 75.1 to 81.1 s wall per clip (timing.json of tests 090-099 and 190-194).
Peak memoryPeak memory of the LTX process not measured. The lowest MemAvailable the queue logged during the 15-clip job was 33 GiB of 121.7 GiB, with a voice model of about 19 GiB and a second ComfyUI also resident.
Tested onNov 2, 2026
ReleasedAug 11, 2026
AI hosts

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

Theo Lindqvist (AI host; designed ElevenLabs voice, AI voice, AI voice)

Narration
Theo Lindqvist (AI voice, ElevenLabs)
Music source
Not supplied
Production
Tests, scoring, scripts and edit by the LEJ Video Lab on a DGX Spark, Nov 2 2026. Hosts are AI characters. All clips are AI-generated with LTX 2.5 22B distilled; no music bed was added to the film.

What it does well

  • One camera sentence in Lightricks' wording produced a real orbit at both seeds (tests 191, 096); the same prompt without it left the camera parked (190, 095)
  • Ending the camera sentence on the final frame can land: at seed 1234 the sun ended exactly behind the tower, as written (096); at seed 42 it ended beside it (191)
  • "Slow dolly in", a film term not on Lightricks' published list, worked at both seeds (tests 091, 098)
  • A described soundtrack replaces the model's default bed: surf and wind instead of held notes for the lighthouse (194, 097), near-silent room tone for the kitchen (094, 099)
  • Counted sounds can land: "a spoon clinking once" gave exactly one transient, at about 3.4 s, at both seeds (094, 099)
  • Prompt length costs nothing in speed: 14.9 to 16.1 s of compute per second of video for every prompt from 16 to 249 words

Where it breaks

  • The negative box does nothing on this graph: swapping the 20-term default negative for a different 12-term list left 121 of 121 frames identical (test 192 vs 191)
  • Over-long prompts (228 and 249 words with quality tags) flipped both scenes to a saturated CG look, rushed the camera through several moves and changed the coffee mid-shot (tests 193, 093)
  • Both over-long prompts came out with loud music (-5.2 and -8.2 LUFS, peaking at or near full scale), probably from one buried mood phrase each; that phrase was not tested on its own (193, 093)
  • Say nothing about sound and the model picks: invented English-sounding speech for a talking host (011), a beat for a cartoon fox (017), a held-note bed for the lighthouse (191, 096)
  • Named sound sources show up on screen: the fridge, wall clock and saucer from the sound sentence appeared in the picture at both seeds (094, 099)
  • Some named sounds never arrive: no gull calls and no clock ticking are visible in any spectrogram
  • Generated loudness is all over the place: -39.1 to -7.9 LUFS across the ten suite prompts, and a described soundtrack came out 27 LU quieter than the undescribed one (094 vs 091), so normalize every clip
  • End states are seed-dependent: the mug dolly reached the rim at seed 42 and stopped at half the frame at seed 1234 (091, 098)

What do the words in an LTX 2.5 prompt actually change? We kept the model, the graph and the seed fixed and changed only the words, five ways, on two scenes.

One camera sentence in Lightricks' wording turned a parked shot into a real orbit at both seeds. The negative box changed zero pixels. The over-long prompts changed everything, mostly for the worse. And one sentence about sound changed the sound and, in the kitchen, the picture too.

How we tested

One model, one graph, one seed, and only the words change. We took two scenes from the lab's fixed prompt suite (P03, a lighthouse on a cliff, and P02, a black coffee mug) and wrote each one five ways:

  1. Plain: one sentence, subject and setting, nothing about the camera, the light style or the sound.
  2. + camera: the plain sentence plus one camera sentence. For the lighthouse we used Lightricks' published wording ("The camera circles around the lighthouse slowly...", from their camera movement guide). For the mug we used a term that is not on that list, "slow dolly in", to see whether a common film word gets through.
  3. Negatives, two ways. Lighthouse: the camera prompt word for word, with the negative box swapped from the lab's 20-term default to a 12-term list ("static camera, still frame, no camera movement, frozen, low quality, blurry, text, watermark, letterbox bars, music, voices, people"). Mug: the camera prompt plus "No people, no hands, no text, no logo on the mug, no camera shake." written into the positive prompt, where people usually put them when they learn the negative box does nothing.
  4. Over-long: 249 words (lighthouse) and 228 words (mug) of the kind of prompt that gets pasted around: stacked adjectives, "8K, masterpiece, trending on artstation", three light sources at once, a crowd of extra subjects and five or six camera moves in a row. Lightricks' README says to keep prompts "within 200 words".
  5. + audio: the camera prompt plus one sentence describing the sound, ending in "No music, no voices."

Everything ran through the lab queue on the DGX Spark on Nov 2, 2026: LTX 2.5 22B distilled, text-to-video, 1280x704, 121 frames at 24 fps, seed 42. The two comparisons this post leans on hardest (plain vs camera on the lighthouse, sound described vs not on both scenes) were run again at seed 1234. That is 15 new clips (tests 090-099 and 190-194). For reference we also used the lab's suite renders of the same two scenes (tests 012 and 013, same seed and settings, the suite's own structured prompts) and the audio of the other eight suite prompts (011-020), none of which say anything about sound.

How the clips were judged:

  • Picture: the lab's 1-5 rubric (prompt adherence, motion, consistency, artifacts, usable), scored from a 12-frame contact sheet per clip plus extra frame grabs where a call was close. One reviewer, one seed pair at most, so treat a one-point difference as noise.
  • How much the frame moved: the structural similarity (SSIM) between the frame at 0.1 s and the frame at 4.9 s. 1.0 means the two frames are identical; a locked camera on a calm scene stays high, a real camera move drops it. It is a proxy, not a camera tracker, and waves or steam moving also lower it.
  • Negatives in the box: a frame-by-frame MD5 of the decoded video and audio of test 192 against test 191.
  • Sound: loudness (EBU R128 integrated LUFS and peak), a spectrogram per clip read by eye, and faster-whisper large-v3 with its voice-activity filter to find speech. Nobody listened: the reviewer here is a language model reading pictures and numbers, so the sound claims rest on measurements, not ears.

The ablation

TestClipScores: adherence / motion / consistency / artifacts / usableFrame change (SSIM) · loudness (LUFS)Speech (VAD)
190lighthouse, Plain, seed 42, 17 words5 / 4 / 5 / 5 / 40.73 · -12.2none
095lighthouse, Plain, seed 1234, 17 words5 / 4 / 5 / 5 / 40.64 · -13.7none
191lighthouse, + camera sentence, seed 42, 40 words4 / 4 / 5 / 4 / 40.68 · -12.2none
096lighthouse, + camera sentence, seed 1234, 40 words5 / 5 / 5 / 5 / 50.75 · -13.7none
192lighthouse, + 12-term negative prompt (negative box), seed 42, 40 words4 / 4 / 5 / 4 / 40.68 · -12.2none
193lighthouse, Over-long (249 words), seed 42, 249 words3 / 3 / 4 / 4 / 30.55 · -5.2none
194lighthouse, + sound sentence, seed 42, 68 words4 / 4 / 5 / 4 / 40.7 · -18.1none
097lighthouse, + sound sentence, seed 1234, 68 words5 / 5 / 5 / 5 / 50.77 · -18.3none
090mug, Plain, seed 42, 16 words5 / 4 / 5 / 5 / 40.85 · -20.4none
091mug, + "slow dolly in", seed 42, 35 words5 / 5 / 5 / 5 / 50.74 · -22.4none
098mug, + "slow dolly in", seed 1234, 35 words4 / 5 / 5 / 5 / 40.74 · -30.3none
092mug, + "No people, no hands, no text...", seed 42, 49 words5 / 5 / 5 / 5 / 50.67 · -27.8none
093mug, Over-long (228 words), seed 42, 228 words3 / 3 / 3 / 3 / 20.66 · -8.2none
094mug, + sound sentence, seed 42, 64 words4 / 5 / 5 / 5 / 40.82 · -49.8none
099mug, + sound sentence, seed 1234, 64 words4 / 5 / 5 / 5 / 40.81 · -41.9none

Rubric scores are 1 to 5 from each test's verdict.md. Words is the positive prompt's word count. Loudness is the clip's own generated soundtrack. Speech is what the VAD-filtered Whisper found ("none" means no speech-like audio at all).

A note on the frame-change column: it didn't do the job we hoped. Breaking surf moves a frame as much as a slow orbit does, so test 095 (a locked camera on big waves) scores lower than test 096 (a full orbit). The camera calls below come from the contact sheets, not that number.

Camera moves: say it in their words, once

The plain lighthouse prompt (17 words, nothing about the camera) gave a parked camera at both seeds: a little drift in test 190, essentially locked in test 095. Add one sentence, "The camera circles around the lighthouse slowly from left to right, keeping the tower centered, and ends with the sun glowing behind it," and the camera moves: a partial orbit at seed 42 (test 191). At seed 1234 (test 096) it's a full, clean orbit, and the sun travels from the left edge of the frame to directly behind the tower on the last frame.

That last part, saying what the shot looks like when the move is done, is the instruction Lightricks' camera guide calls the most valuable one in its list. At seed 1234 the model hit it exactly.

Compare the lab's suite prompt for the same scene (test 013): "Aerial drone shot orbiting slowly around a lighthouse... Smooth steady orbit." Same seed, same settings, and it gave a slight drift, no orbit. Different words around it, so it's not a clean comparison, but it's the same pattern our earlier model review saw with big moves. The suite prompt is shorter than ours (33 words against 40); what differs is the verb ("orbiting" against the guide's "circles around") and that ours says where the move ends.

The camera guide says to use its terms exactly and calls synonyms guesses. For the mug we used one that isn't on its list, "slow dolly in", and it worked at both seeds: a full push to the rim in test 091 (5 out of 5 on every criterion) and a shorter one in test 098 that stopped at about half the frame. So a standard film word can get through. The end state ("until its rim and the steam fill the frame") is what one seed honored and the other didn't.

Negative prompts: the box does nothing, the words in the prompt do something

Test 192 is test 191 word for word, with a different negative. Test 191 carried the lab's 20-term default negative ("black borders, letterbox bars, ... static, frozen face", listed under Settings). Test 192 swapped it for 12 terms: "static camera, still frame, no camera movement, frozen, low quality, blurry, text, watermark, letterbox bars, music, voices, people".

Seven of those are new, among them "static camera" and "music". If the negative box did anything, those should have pushed the shot toward more motion and changed the sound.

All 121 decoded video frames are bit-identical to test 191. Same MD5, every frame. The audio isn't bit-identical, but it's nearly so: a correlation of 0.99985, with the difference 35 dB below the signal.

The negative is never evaluated at CFG 1.0 (see Setup), so that small difference most likely comes from the audio decode or encode step. We haven't isolated it.

So the negative box on this graph is decoration. The tempting workaround is to move the "no" list into the prompt itself. Test 092 did that for the mug: the camera prompt plus "No people, no hands, no text, no logo on the mug, no camera shake." None of those things showed up, and the dolly in was clean (5 out of 5 across the board).

What did change was the whole composition: a closer start, a different chair, a different window, from the same seed. Every extra word reshuffles the frame. It didn't backfire here, on one seed, but it spent words on things that weren't going to happen anyway. Lightricks' advice to describe what you want, in the positive, is the cheaper habit.

Over-long prompts

Both over-long prompts (249 words for the lighthouse, 228 for the mug) flipped the look. The scenes stopped being photoreal and turned into a saturated, game-engine style render: a red-and-white striped tower with red boats and gulls (test 193), a kitchen with a croissant, a succulent, glasses, a cat, candles and a blue neon sign (test 093). The quality-tag pile ("8K, HDR, masterpiece, trending on artstation, hyperrealistic") produced the least realistic clips of the night.

The camera tried to do everything at once: a fast low pass with heavy motion smear, a push in, a rising pull-back, all inside five seconds, with no double orbit and no pan. In the mug clip the coffee changes mid-shot from dark to latte art. And both clips came out with music, and loud, probably from the one mood phrase buried in each prompt ("epic orchestral mood", "cozy lofi music mood"); we didn't test that phrase on its own, and the fox and the plain lighthouse show the model writes music unasked too: -5.2 LUFS for the lighthouse (the loudest clip we measured) and -8.2 for the mug, both peaking at or within 0.3 dB of full scale.

Scores: 3 out of 5 or lower on adherence, motion and usability for both. Lightricks' README asks for prompts "within 200 words". These were 228 and 249 words, and each one packed five or six camera moves, three kinds of light and a crowd of extra objects into five seconds of video.

What the model hears

LTX 2.5 doesn't make a silent clip and add sound later. Picture and audio come out of the same 22B transformer in one pass, from the same prompt. So every word you write is also an instruction to the sound, and every word you leave out is a decision the model makes for you.

Here is what it decided on its own across the lab's suite, where no prompt mentions sound, next to our ablation clips with and without a sound sentence:

TestWhat the prompt says about soundLoudness (LUFS) · peak (dBFS)Speech (VAD)Spectrogram read
011nothing ("talks to the camera")-17.5 · -1.6yes: "i like freedom of the clap shims and empty fashion who has t"voiced harmonics in word-length bursts (speech); Whisper heard English words that make no sense
014nothing-7.9 · 0.0noneloud broadband rumble with clicks (wheels, board); peaks at 0.0 dBFS
015nothing-39.0 · -24.8nonenot read (loudness only)
016nothing-39.1 · -22.3nonenot read (loudness only)
017nothing-12.3 · -0.0noneevenly spaced beats ~0.42 s apart with tonal notes on top (music)
018nothing-32.9 · -22.0nonenot read (loudness only)
012nothing-29.0 · -17.6nonequiet; energy mostly below ~1 kHz
013nothing-11.8 · -2.9nonelow broadband rumble below ~600 Hz with faint steady tones 1.7-4 kHz
019nothing-22.5 · -11.4nonenot read (loudness only)
020nothing-19.5 · -8.8nonenot read (loudness only)
191nothing-12.2 · -2.7noneenergy below ~1 kHz in steady tonal bands that change in steps (held-note, music-like bed); little above 2 kHz
194surf, wind, distant gulls; "No music, no voices."-18.1 · -6.0nonebroadband noise up to ~8 kHz (surf/wind); no stepped tones; no gull-call shapes visible
096nothing-13.7 · -0.9nonestacked steady bands below ~1.5 kHz, little above 2 kHz (same pattern as 191)
097surf, wind, distant gulls; "No music, no voices."-18.3 · -5.9nonebroadband noise up to ~8 kHz (surf/wind), no stacked steady bands; gull calls not visible
091nothing-22.4 · -6.9nonesteady low bands below ~500 Hz (hum/pad-like), stepping early in the clip
094room tone, fridge hum, clock ticking, one spoon clink off screen; "No music, no voices."-49.8 · -38.0nonevery quiet noise floor below ~400 Hz (room tone) and one short transient at ~3.4 s up to ~1.7 kHz; no regular ticking visible
098nothing-30.3 · -11.4nonesteady tonal bands below ~2 kHz (hum/pad-like), louder in the first second
099room tone, fridge hum, clock ticking, one spoon clink off screen; "No music, no voices."-41.9 · -30.9nonequiet room tone below ~400 Hz and one sharp transient at ~3.4 s up to ~6 kHz (one clink, as asked); no regular ticking visible

What the suite says: say nothing about sound, and the model fills it with what the picture implies. The talking host got speech, and Whisper transcribed English-shaped words that mean nothing ("i like freedom of the clap shims and empty fashion..."). The cartoon fox got a beat, evenly spaced hits about 0.42 s apart with notes on top.

The skateboard got rumble and clicks peaking at full scale. Across the ten suite prompts the integrated loudness ran from -39.1 to -7.9 LUFS: a 31 LU spread from a model that sets its own levels.

What one sound sentence did, at two seeds each:

  • Lighthouse. With nothing about sound (tests 191 and 096), the soundtrack was a steady, held-note bed under 1.5 kHz, more music than sea. With "the deep roar of waves crashing on the rocks below, a steady wind gusting across the cliff, and gulls calling in the distance. No music, no voices." (tests 194 and 097), the held notes went away and broadband surf and wind noise up to about 8 kHz took over, 4.6 to 5.9 LU quieter. No gull calls are visible in either spectrogram.
  • Mug. With nothing about sound (091, 098), a low hum-like bed. With room tone, a fridge hum, a ticking clock and "a spoon clinking once against a saucer off screen" (094, 099), the clip dropped to near silence (-49.8 and -41.9 LUFS) with one sharp transient at about 3.4 s in both seeds. One clink, as asked. No regular ticking is visible.

And the part we didn't expect: in the kitchen, the sound sentence changed the picture. Both seeds of the sound-described mug grew a refrigerator, a wall clock and a saucer under the mug, none of which the picture part of the prompt mentions. The spoon was supposed to be off screen; the saucer showed up anyway.

The 22B model reads the whole prompt for both outputs. If you name a sound source, expect to see it unless the shot can't contain it.

Practical reading: write the sound sentence every time, because the default is a guess (and sometimes music or fake speech). Name the sources, end with "No music, no voices." (a third-party convention; Lightricks' own audio article says unnamed layers are "left to the model's defaults"), and plan to normalize: the described soundtracks came out much quieter, one of them 27 LU quieter than the undescribed take.

The prompt we'd write now

Three sentences, in this order, and leave the negative box alone:

  1. The shot in plain words. Subject, place, light. "A white lighthouse on a rocky coastal cliff at golden hour, waves breaking against the rocks below."
  2. One camera sentence, ending on what the frame looks like when the move is done. "The camera circles around the lighthouse slowly from left to right, keeping the tower centered, and ends with the sun glowing behind it." Use the guide's wording when you can; "slow dolly in" worked too.
  3. One sound sentence naming each source, then "No music, no voices." And only name sources you're happy to see in frame.

That's test 097: 68 words, a clean orbit that lands its end state, surf and wind instead of a music bed, and 5 out of 5 on every picture criterion. Skip the quality tags, skip the "no" list, and stop at one camera move.

What we'd test next

  • More seeds. Two seeds is enough to see a pattern, not enough to call a rate.
  • Whether "off screen" can be made to stick (a closed-door kitchen, say) or whether every named sound source ends up in frame.
  • A repeat of test 191 with nothing changed, to pin the tiny audio difference in 192 on the decoder rather than leave it as "most likely".
  • The same ablation on image-to-video, where the start frame already fixes the look and the camera sentence may matter more.

Setup and gotchas

  • The graph is the stock two-stage LTX 2.5 graph, built by our script. Stage one runs 8 steps at half size (640x352) on a fixed sigma list, a 2x latent upscaler takes it to 1280x704, and stage two runs 3 more steps. Sampler euler_ancestral, LTXVDualCFGGuider at video_cfg 1.0 and audio_cfg 1.0, 24 fps, 121 frames (frame counts have to be 8k+1). The exact API graph for test 191 is linked at the bottom of this page.
  • Your words go through a language model before the video model sees them. The text encoder is a Gemma 4 12B fine-tuned by Lightricks with a trained projection (gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot, 15.4 GB on disk). Lightricks says stock Gemma 4 is not a drop-in substitute (LTX-2 README). The video model only ever gets that encoder's output, which is why literal, ordered prose works better than a tag list.
  • The negative prompt is never read at CFG 1.0. With both guider scales equal, ComfyUI's Guider_LTXAVDualCFG falls back to plain CFG (comfy_extras/nodes_lt.py), and plain CFG at a scale of 1.0 skips the unconditional pass entirely (comfy/samplers.py line 610 in ComfyUI 0.43.0: if math.isclose(cond_scale, 1.0) ... uncond_ = None). The negative text still gets encoded, then nothing uses it. The distilled checkpoint is made for "a fixed 8-step schedule, CFG=1" (model card), so raising CFG to make the negative count is off-label. The same wiring is reported for the stock LTX 2.3 Comfy templates in Comfy-Org/workflow_templates #919, and the 2.5 templates carry it too.
  • Every clip has a soundtrack whether you ask or not. LTX 2.5 generates audio jointly with the picture, and the graph decodes it into the mp4 (AAC, 48 kHz stereo). If the prompt says nothing about sound, the model picks something.
  • Whisper invents words on noise. Run without its voice-activity filter on the suite's lighthouse and mug clips, faster-whisper large-v3 transcribed "Thanks for watching!" for both. With the filter on it returned nothing for every clip except the one with a voice. Every speech claim on this page uses the VAD-filtered run, checked against the spectrogram. Positive control: the same setup transcribed an AI host's 22 s intro word for word.
  • Same seed, same settings, one change at a time. Every variant of a scene used seed 42 at 1280x704; the camera and audio comparisons were repeated at seed 1234. The stage-two noise seed is fixed at 42 in our graph for every test.

Settings

SettingValue
Transformerltx-2.5-22b-distilled-transformer-comfy-int8-convrot
Text encodergemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot
VAEsltx-2.5-video-vae-bf16, ltx-2.5-audio-vae-bf16
Upscalerltx-2.5-latent-spatial-upscaler-x2-bf16-1.0
Stage 1640x352, sigmas 1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0
Stage 21280x704 after the 2x latent upscale, sigmas 0.85, 0.725, 0.4219, 0.0
Samplereuler_ancestral, both stages
GuidanceLTXVDualCFGGuider, video_cfg 1.0, audio_cfg 1.0
Frames / fps121 frames at 24 fps (5.04 s)
Seeds42 (all variants), 1234 (camera and audio repeats); stage-two seed 42
Negativeevery test except 192: the lab's 20-term default "black borders, letterbox bars, pillarbox, text, letters, words, writing, watermark, logo, subtitles, low quality, blurry, flicker, warping, morphing, melting, deformed, extra limbs, static, frozen face" (inert, kept for parity with the rest of the lab); 192: the 12-term list in its prompt field
DecodeVAEDecodeTiled 512 / 64 / 64 / 16
ComfyUI0.43.0, commit 426f5ee, PyTorch 2.11.0+cu130

Speed and memory

All 15 clips ran back to back in one queue job on Nov 2 (13:49 to 14:08 ET, 1170 s wall, logs/gpu-queue.log). Per clip: 75.1 to 81.1 s of wall time for 5.04 s of video, so 14.9 to 16.1 s of compute per second of video (median 15.5), from each test's timing.json. Prompt length made no measurable difference to speed: the 249-word prompt (193) took 78.1 s and the 17-word one (190) took 81.1 s. The first clip followed LTX host renders on the same server (same weights; cold load not measured) and took 75.1 s.

Memory: the lowest MemAvailable the queue logged during the job was 33 GiB of the box's 121.7 GiB (the LTX server, a voice model of about 19 GiB and a second ComfyUI were all resident). Peak memory of the LTX process alone was not measured.

Sources

Every clip on this page is AI-generated. The hosts in the videos, Daisy Lane and Theo Lindqvist, are AI characters with AI voices.

ComfyUI API graph for test 191 (lighthouse, camera sentence); every test used this graph with its own prompt and seed (and, for 192, its own negative) ↗