We asked LTX 2.5 for servo sounds on a silent robot clip. It gave the robot a voice.

Rated G

LTX 2.5 will write a soundtrack for a clip you already have in about 25 seconds on a DGX Spark. What it writes depends on what it sees: an engine for the car, a low hum for the beach, and speech every time the robot's mouth opens. The "no voices" negative prompt does nothing until you raise one number in the graph. Raise it and the seed decides: the same words came back clean on one seed and still talking on another.

The walkthrough (191 s): AI hosts Daisy Lane and Rafael Duarte (AI voices, ElevenLabs) play the sound LTX 2.5 wrote for the car, the beach and the robot, show why the robot talks, and the setting and seed roll that mostly fix it. The clips play level matched here; the real level differences are in the captions below. Lip-sync is loose at pauses, so the edit cuts to the clips over the long ones.
Prompt and settings

Settings: Four host segments (51.0 s of speech, eleven_v4) made with LTX 2.5 image+audio-to-video from each host's on-set still, seed 42; eight voice-over panels, five listen blocks and cut-aways in Remotion using this post's own clips; 1920x1080, 30 fps

Reference images (2)
Roadster, seed 42. One continuous engine-and-wind track. It's about 10 dB louder than the beach clip, so mind your volume.
Prompt and settings

Prompt: a small sports car engine humming at speed on an open coastal road, wind rushing past, distant ocean surf, no music

Settings: LTX 2.5 22B distilled, int8 transformer, sound only: the picture is encoded at 640x352 and frozen, and LTX samples just the audio. 121 frames at 24 fps (5.04 s), 48 kHz stereo out. euler_ancestral, 8 steps, CFG 1.0 for video and audio. Seed 42, 51.4 s (the first job after a model load). Negative prompt (inert at CFG 1): speech, talking, voices, singing, vocals, laughter, music, distorted, noise, hum. The picture is an LTX 2.5 image-to-video take from the lab's short film, and this pass replaced its sound. Rendered Nov 2 2026. The two people are AI-generated characters.

Reference images (1)
Beach, seed 42. We asked for waves, a breeze and gulls. It's a steady low hum, and the two people laughing got no voice.
Prompt and settings

Prompt: gentle ocean waves rolling onto sand, soft breeze, distant seagulls, no music, no voices

Settings: LTX 2.5 22B distilled, int8 transformer, sound only: the picture is encoded at 640x352 and frozen, and LTX samples just the audio. 121 frames at 24 fps (5.04 s), 48 kHz stereo out. euler_ancestral, 8 steps, CFG 1.0 for video and audio. Seed 42, 24.0 s. Negative prompt (inert at CFG 1): speech, talking, voices, singing, vocals, laughter, music, distorted, noise, hum. The picture is an LTX 2.5 image-to-video take from the lab's short film, and this pass replaced its sound. Rendered Nov 2 2026. The two people are AI-generated characters.

Reference images (2)

The same robot clip with four soundtracks: switch the audio

  • Clip 2: Robot, first prompt, seed 3. We asked for room tone, a servo whirr and plastic clicks. Listen at 0, 1 and 3.5 seconds, each time the mouth opens.
    Prompt and settings

    Prompt: Quiet cozy home office ambience: soft room tone, a gentle mechanical servo whirr and tiny plastic clicks as a small robot moves its head and waves its arm, a faint hum from a laptop. No speech, no music.

    Settings: LTX 2.5 22B distilled, int8 transformer, sound only: the picture is encoded at 640x352 and frozen, and LTX samples just the audio. 121 frames at 24 fps (5.04 s), 48 kHz stereo out. euler_ancestral, 8 steps, CFG 1.0 for video and audio. Seed 3, 27.1 s. Negative prompt (inert at CFG 1): speech, talking, voices, music, singing, distorted, noise, hum. The picture is a silent Wan 2.2 image-to-video clip, copied to 24 fps. Rendered Nov 1 2026. The frame is cropped about 10% to cut a logo off the laptop; the sound is untouched.

    Reference images (1)
  • Clip 3: Robot, second prompt ("Sound effects only, no people and no voices"), seed 11. Same three moments. The first and last are noise now; the middle one still trips a speech detector.
    Prompt and settings

    Prompt: Sound effects only, no people and no voices. Soft room tone of a quiet office, the small electric whirr of a toy robot servo motor as its head turns, light plastic clicks as its arm waves, a faint laptop fan hum.

    Settings: LTX 2.5 22B distilled, int8 transformer, sound only: the picture is encoded at 640x352 and frozen, and LTX samples just the audio. 121 frames at 24 fps (5.04 s), 48 kHz stereo out. euler_ancestral, 8 steps, CFG 1.0 for video and audio. Seed 11. Negative prompt (inert at CFG 1): speech, talking, human voice, voices, words, dialogue, singing, vocals, whisper, mumbling, breathing, music, melody. The picture is a silent Wan 2.2 image-to-video clip, copied to 24 fps. Rendered Nov 1 2026. The frame is cropped about 10% to cut a logo off the laptop; the sound is untouched.

    Reference images (1)
  • Clip 6: Robot, second prompt, seed 11, with audio_cfg raised from 1 to 3. Same three moments, and a speech detector finds no voice in them now. It's about 10 dB louder than the same take at 1, so mind your volume.
    Prompt and settings

    Prompt: Sound effects only, no people and no voices. Soft room tone of a quiet office, the small electric whirr of a toy robot servo motor as its head turns, light plastic clicks as its arm waves, a faint laptop fan hum.

    Settings: LTX 2.5 22B distilled, int8 transformer, sound only: the picture is encoded at 640x352 and frozen, and LTX samples just the audio. 121 frames at 24 fps (5.04 s), 48 kHz stereo out. euler_ancestral, 8 steps, CFG 1.0 for video, audio_cfg 3.0. Seed 11, 40.2 s. Negative prompt: speech, talking, human voice, voices, words, dialogue, singing, vocals, whisper, mumbling, breathing, music, melody. The picture is a silent Wan 2.2 image-to-video clip, copied to 24 fps. Rendered Nov 8 2026. The frame is cropped about 10% to cut a logo off the laptop; the sound is untouched.

    Reference images (1)
  • Clip 7: Robot, first prompt, seed 3, with audio_cfg at 2. More voice than the same take at 1, not less: a speech detector flags all three bursts now.
    Prompt and settings

    Prompt: Quiet cozy home office ambience: soft room tone, a gentle mechanical servo whirr and tiny plastic clicks as a small robot moves its head and waves its arm, a faint hum from a laptop. No speech, no music.

    Settings: LTX 2.5 22B distilled, int8 transformer, sound only: the picture is encoded at 640x352 and frozen, and LTX samples just the audio. 121 frames at 24 fps (5.04 s), 48 kHz stereo out. euler_ancestral, 8 steps, CFG 1.0 for video, audio_cfg 2.0. Seed 3, 40.1 s. Negative prompt: speech, talking, voices, music, singing, distorted, noise, hum. The picture is a silent Wan 2.2 image-to-video clip, copied to 24 fps. Rendered Nov 8 2026. The frame is cropped about 10% to cut a logo off the laptop; the sound is untouched.

    Reference images (2)

One prompt, one negative, audio_cfg raised: seed 11 at 2 and 3, seed 3 at 3

  • Clip 8: Robot, second prompt, seed 11, audio_cfg 2. No voice for the speech detector here either, and about 2 dB quieter than the same take at 3.
    Prompt and settings

    Prompt: Sound effects only, no people and no voices. Soft room tone of a quiet office, the small electric whirr of a toy robot servo motor as its head turns, light plastic clicks as its arm waves, a faint laptop fan hum.

    Settings: LTX 2.5 22B distilled, int8 transformer, sound only: the picture is encoded at 640x352 and frozen, and LTX samples just the audio. 121 frames at 24 fps (5.04 s), 48 kHz stereo out. euler_ancestral, 8 steps, CFG 1.0 for video, audio_cfg 2.0. Seed 11, 30.1 s. Negative prompt: speech, talking, human voice, voices, words, dialogue, singing, vocals, whisper, mumbling, breathing, music, melody. The picture is a silent Wan 2.2 image-to-video clip, copied to 24 fps. Rendered Nov 10 2026. The frame is cropped about 10% to cut a logo off the laptop; the sound is untouched.

    Reference images (1)
  • Clip 9: Robot, seed 3, given the second prompt and its long negative at audio_cfg 3: the words that quieted seed 11. Seed 3 talks anyway, in the same three bursts. It's the loudest track here, so mind your volume.
    Prompt and settings

    Prompt: Sound effects only, no people and no voices. Soft room tone of a quiet office, the small electric whirr of a toy robot servo motor as its head turns, light plastic clicks as its arm waves, a faint laptop fan hum.

    Settings: LTX 2.5 22B distilled, int8 transformer, sound only: the picture is encoded at 640x352 and frozen, and LTX samples just the audio. 121 frames at 24 fps (5.04 s), 48 kHz stereo out. euler_ancestral, 8 steps, CFG 1.0 for video, audio_cfg 3.0. Seed 3, 64.8 s (the first job after a model load). Negative prompt: speech, talking, human voice, voices, words, dialogue, singing, vocals, whisper, mumbling, breathing, music, melody. The picture is a silent Wan 2.2 image-to-video clip, copied to 24 fps. Rendered Nov 10 2026. The frame is cropped about 10% to cut a logo off the laptop; the sound is untouched.

    Reference images (1)

Verdict

Use LTX 2.5's sound-only pass when the shot has one loud, steady thing in it, like an engine, and pull the result down about 10 dB. Skip it for quiet scenes and for surf. Our robot's moving mouth got a voice, and the negative prompt can't stop that at CFG 1, where this graph runs. Raise audio_cfg to 2 or 3 with a long "no voices" negative and roll seeds: in our test the seed decided, not the words. It needs nothing you don't already have for LTX video.

Credits and settings

WorkflowLTX 2.5 audio branch as Foley (sound for a silent clip)
MakerLightricks
Versionltx-2.5-22b-distilled-transformer-comfy-int8-convrot with ltx-2.5-audio-vae-bf16; audio-only sampling, video latent frozen
What it doesvideo-to-audio (Foley)
LicenseLTX-2.x Community License Agreement
HardwareDGX Spark: NVIDIA GB10, ARM (aarch64), 128 GB unified memory. ComfyUI 0.43 core nodes, int8 LTX 2.5 transformer and Gemma 4 12B text encoder, bf16 video and audio VAEs.
Speed5.2 s of compute per second of video. Wall time for the sound pass alone, with the models already loaded. About 6 to 8 with audio_cfg above 1.
Peak memorySame files as LTX 2.5 text-to-video, whose ComfyUI process peaked at 42,254 MiB in our Nov 7 test (nvidia-smi, per process). Plan for that.
Tested onNov 10, 2026
ReleasedAug 11, 2026
AI hosts

Daisy Lane (AI host; AI voice)

Rafael Duarte (AI host; AI voice)

Narration
Not supplied
Music source
Not supplied

What it does well

  • Fast. The sound pass takes about a third of the time LTX needs to make picture and sound together.
  • Sound lands where the picture moves. Two seeds and two prompts put the robot's three bursts at the same three moments.
  • One loud, steady source works: both car shots got a continuous engine-and-wind track.
  • Core ComfyUI 0.43 nodes and the LTX 2.5 files you already have. LTX never redraws your frames.

Where it breaks

  • A moving mouth can get a voice. Asked for servo whirrs with 'No speech, no music', the robot clip came back with 1.5 s of speech.
  • Dead air between events. More than half of the robot track is silence, where we asked for room tone.
  • Surf comes out as a low hum with nothing above 2.5 kHz.
  • Car tracks come out about 10 dB hotter than the beach, with samples at full scale.
  • The negative prompt is inert at CFG 1.0, which is where this graph runs.
  • Raising audio_cfg doesn't fix every seed. With the same prompt and negative at 3, seed 11 came back clean and seed 3 still talked.
  • Takes with audio_cfg raised came back 7 to 10 dB louder than at 1.

Wan 2.2 hands you a silent clip. LTX 2.5 is probably loaded on the same box already, and it writes sound whenever it makes its own video. So the obvious move is to hold your picture still and let LTX write only the sound.

It will. What it writes depends on what it sees, a lot more than on what you ask for.

Watch the robot's mouth

We asked for room tone, a servo whirr and plastic clicks, with "No speech, no music" right in the prompt. Whisper, asked to write down what LTX made instead, hears "Wow, they made cool side." and "High time."

The robot's mouth is open at the start of the clip, again around 1 to 1.5 seconds, and again at 3.5 seconds, when it waves. The sound comes in three bursts, at those three moments, and more than half of the clip between them is dead air. Our read: LTX saw a mouth move and gave it something to say.

A new seed and a blunter prompt landed in the same three places

The second try used a different seed and a prompt that opens with "Sound effects only, no people and no voices." The bursts didn't move: 0 to 0.5 s, 0.75 to 2 s, 3.5 to 4.25 s.

What changed is what fills them. The first and last turned into noise, and a speech detector's count dropped from 1.5 seconds of voice to 0.45. The middle burst is what it still catches. Those two tries are the first two clips in the comparison above, so switch the audio between them and hear it.

So the picture decides when there's sound. The prompt and the seed only get a vote on what kind.

"No voices" in the negative prompt does nothing at CFG 1

Both tries had speech and voices in the negative prompt. That never had a chance. This graph runs LTX 2.5 distilled at CFG 1.0 for video and for audio, and at 1.0 the negative prompt isn't part of the math.

Raise audio_cfg, then roll seeds

One node sets it: LTXVDualCFGGuider, which has video_cfg and audio_cfg side by side. We left the video at 1, raised the audio, and reran both tries with nothing else changed.

The second try lost its voice at 2 and at 3. Same three bursts, and the speech detector finds nothing in any of them. The middle burst, the one the detector caught before, now holds a steady pitch where it used to slide around like a voice. Play it and decide if that's a servo. It also came back louder: about 7 dB at 2 and 10 at 3.

The first try went the other way. At 2 it has more voice than it started with: the detector flags all three bursts, 2 seconds of the 5. At 3 it's silent from start to finish.

So was it the seed that cleaned up the second try, or its blunter prompt and longer negative? We gave seed 3 the second try's prompt and negative at 3. It talked: the same three bursts, 2 seconds of voice, and the loudest track on this page, peaking just under full scale. The words that quieted seed 11 did nothing for seed 3. The seed decides.

All three takes on the second prompt sit in the second comparison above. On your own clip, set audio_cfg to 2 or 3 with a long "no voices" negative, roll a handful of seeds, and keep one that doesn't talk. A speech detector sorts them in seconds on a CPU. Listen to the keeper before it goes in a cut.

How to run it on your own clip

Core ComfyUI 0.43 nodes, no custom nodes, and the same LTX 2.5 files you already load for video. The workflow file at the bottom of the page is the exact graph from the second robot try, with audio_cfg at 1.

  1. Make a 24 fps copy of the clip: ffmpeg -i clip.mp4 -vf fps=24 -an clip24.mp4. The Wan 2.2 14B models render at 16 fps, and LTX wants 24. The copy runs the same length, so the sound will fit your original file too.
  2. Load the copy and keep 8k+1 frames (121 frames is 5 seconds). Scale them to half size, 640x352 for a 1280x704 clip, and encode them with the LTX video VAE.
  3. LTXVFreezeLatent on that latent, LTXVEmptyLatentAudio for the sound, LTXVConcatAVLatent to join the two.
  4. Sample 8 steps with euler_ancestral and the distilled first-stage sigmas. Only the audio moves.
  5. LTXVSeparateAVLatent, decode with the LTX audio VAE, and mux the sound onto the full-size frames. LTX never redraws them.
  6. Put the sound on your original clip and turn it down in the same pass: ffmpeg -i original.mp4 -i foley.mp4 -map 0:v -map 1:a -c:v copy -af volume=-10dB final.mp4. Your picture is copied untouched. Leave out -af volume=-10dB when the level is already fine.
  7. If a voice shows up, change audio_cfg on the LTXVDualCFGGuider node from 1 to 2 or 3, leave video_cfg alone, and try new seeds until one comes back without it.

The car works, loud. The beach hums.

The car is the easy case. Both versions of the shot came back as one continuous engine-and-wind track, and the speech detector found no voice in either, with two people in plain view in the roadster. They also came back about 10 dB hotter than the beach, with samples hitting the ceiling. That's what step 6 is for.

The beach asked for waves, a breeze and gulls, and got a steady low hum. On a spectrogram it goes black above 2.5 kHz, where LTX's own text-to-video surf fills the whole picture with noise. Both pictures are in the reference images under the beach clip.

Two people laugh through that beach shot and LTX gave them no voice, so a moving mouth isn't a sure trigger. Check every take.

25 seconds a clip, and the memory of LTX video

Runs with the models loaded took 24 to 29 seconds for a 5-second 1280x704 clip. That's about 5 seconds of compute per second of video, against about 15 when LTX makes the picture too. With audio_cfg above 1 it's 30 to 40 seconds. The first job after a model load took 51 to 89 seconds.

It loads the same files as LTX text-to-video, which peaked at 42,254 MiB for us. Plan for that much.

Here's every clip by loudness. LUFS is a loudness scale, and ten points up is roughly twice as loud.

ClipLoudnessVoice a speech detector found
Robot, first prompt, seed 3 (shown)-21.7 LUFS1.54 s
Robot, first prompt, seed 3, audio_cfg 2 (shown)-15.0 LUFS1.98 s
Robot, first prompt, seed 3, audio_cfg 3-70.0 LUFSnone
Robot, second prompt, seed 11 (shown)-21.1 LUFS0.45 s
Robot, second prompt, seed 11, audio_cfg 2 (shown)-13.8 LUFSnone
Robot, second prompt, seed 11, audio_cfg 3 (shown)-11.5 LUFSnone
Robot, second prompt, seed 3, audio_cfg 3 (shown)-8.0 LUFS2.05 s
Roadster (shown)-10.5 LUFSnone
Car, earlier version of the shot-9.1 LUFSnone
Beach (shown)-20.0 LUFSnone
Beach, earlier version of the shot-19.7 LUFSnone

Use it for engines. Keep it away from mouths.

  • One loud, steady thing on screen (an engine, wind, machinery): run it, pull it down 10 dB, move on.
  • A character whose mouth moves: expect a voice. Roll a few seeds with audio_cfg at 2 or 3 and a long negative, use a sound library, or pick a take where the mouth stays shut.
  • Quiet rooms and small sounds: you get dead air between events. Lay your own room tone under it.
  • Waves, rain, anything that's mostly hiss: let LTX make picture and sound together, or use a recording.

The real fix is a model built for video-to-audio. HunyuanVideo-Foley is the one we'd try first. It isn't on our Spark yet: its official weights ship as pickle files, and we only load safetensors from official repos.

ComfyUI API graph (robot clip, second prompt, audio_cfg 1) ↗