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.
- 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. - 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.
LTXVFreezeLatenton that latent,LTXVEmptyLatentAudiofor the sound,LTXVConcatAVLatentto join the two.- Sample 8 steps with
euler_ancestraland the distilled first-stage sigmas. Only the audio moves. LTXVSeparateAVLatent, decode with the LTX audio VAE, and mux the sound onto the full-size frames. LTX never redraws them.- 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=-10dBwhen the level is already fine. - If a voice shows up, change
audio_cfgon theLTXVDualCFGGuidernode from 1 to 2 or 3, leavevideo_cfgalone, 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.
| Clip | Loudness | Voice a speech detector found |
|---|---|---|
| Robot, first prompt, seed 3 (shown) | -21.7 LUFS | 1.54 s |
Robot, first prompt, seed 3, audio_cfg 2 (shown) | -15.0 LUFS | 1.98 s |
Robot, first prompt, seed 3, audio_cfg 3 | -70.0 LUFS | none |
| Robot, second prompt, seed 11 (shown) | -21.1 LUFS | 0.45 s |
Robot, second prompt, seed 11, audio_cfg 2 (shown) | -13.8 LUFS | none |
Robot, second prompt, seed 11, audio_cfg 3 (shown) | -11.5 LUFS | none |
Robot, second prompt, seed 3, audio_cfg 3 (shown) | -8.0 LUFS | 2.05 s |
| Roadster (shown) | -10.5 LUFS | none |
| Car, earlier version of the shot | -9.1 LUFS | none |
| Beach (shown) | -20.0 LUFS | none |
| Beach, earlier version of the shot | -19.7 LUFS | none |
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_cfgat 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.