Video SFX

Generate synchronized sound effects, foley, or ambient audio for a video clip.

Overview

The Video SFX node takes an input video plus an optional text prompt and produces an mp4 with AI-generated sound effects merged in. It uses Replicate’s zsxkib/mmaudio model (based on the open-source MMAudio project) and is ideal for adding rain, footsteps, ambience, room tone, foley, or other diegetic sounds to silent or AI-generated video.

⚠️ Replaces the existing audio track. The original audio of the input video is dropped and replaced with the generated SFX. To keep the original audio, route this node’s output into a Merge Video & Audio node with the original audio as a second input.

Configuration

Field	Type	Default	Description
Prompt	Textarea	—	Describes the desired sound (e.g., “rain on metal roof”, “footsteps in dry leaves”). Connected text/prompt nodes override the config-panel value.
Negative prompt	Textarea	music	Sounds to exclude. mmaudio’s default keeps the output SFX-only — change it if you actually want musical SFX.
Versions	Number	1	How many takes to generate per run (1–4). Each version uses a distinct seed for variety; cost scales linearly.
cfg_strength	Number	4.5	Advanced — guidance scale (1.0–10.0).
num_steps	Number	25	Advanced — diffusion steps (10–50).
seed	Number	random	Advanced — leave blank or -1 for random. When set, version N uses seed + N.

Inputs & Outputs

Inputs:

• Video (required) — source video (mp4, mov, webm). Maximum 300 seconds.
• Prompt (optional) — text describing the desired SFX. May be wired from a Text Prompt or Generate Text node.
• Negative (optional) — text describing sounds to exclude. Defaults to "music".

Outputs:

• Video — an mp4 with the generated SFX merged in (original audio dropped).

Credit Cost

Duration is derived automatically from the input video via ffprobe. The duration is rounded up into the next bucket; the bucket determines the BASE cost. The system markup (default 25%) is applied on top.

Input video duration	Bucket	BASE credits / version	Visible credits / version
≤ 8s	:8s	1	2
≤ 15s	:15s	1	2
≤ 30s	:30s	2	3
≤ 60s	:60s	3	4
≤ 120s	:120s	5	7
≤ 300s	:300s	11	14

For multi-version runs the cost multiplies: versions × per-version. Worked examples:

Duration	Versions	BASE total	Visible total
5s	1	1	2
8s	1	1	2
12s	1	1	2
30s	1	2	3
31s	1	3	4
60s	4	12	15
180s	1	11	14

If ffprobe fails to derive a duration, the :8s bucket is used as a fallback and a warning is logged.

Constraints

• Supported input formats: MP4 (H.264), MOV, WebM.
• Maximum duration: 300 seconds (5 minutes). Longer videos are rejected with HTTP 400 video_duration_exceeds_limit.
• Maximum versions per run: 4.
• Recommended input resolution: up to 1080p (higher resolutions are untested and increase processing time).

Determinism

• seed defaults to random per run.
• If you set a seed and versions > 1, each version uses seed + index so all takes are reproducible but distinct.
• If you leave seed blank (or set -1), every version uses a fresh random seed.

Best Practices

• Write prompts that describe the sound, not the on-screen action (“crackling fire” beats “campfire scene”).
• Keep clips short for iteration — :8s and :15s buckets share the same 1-credit BASE, so an 8-second test clip costs the same as a 15-second one.
• If you want layered audio (e.g. SFX over music), generate the SFX here then mix the music in via Mix Audio.

Common Use Cases

• Add foley or ambience to silent AI-generated video.
• Generate room tone, weather, or environmental SFX for short cinematic clips.
• Produce multiple alternate takes (versions: 4) and pick the best via a downstream Reduce node.
• Layer with the original dialogue/music using Merge Video & Audio.

Powered by

This node sends your video to Replicate for processing using the zsxkib/mmaudio model, based on the open-source MMAudio project. See Replicate’s terms for licensing and attribution.