KIE • Sound Cue Sheet

Every tool in the "AI sound design" space has the same failure mode: it watches your video, decides a car door closed at 4 seconds, and generates a car door. Sometimes it's right. When it's wrong you've paid for a sound effect that doesn't exist in your footage.

KIE • Sound Cue Sheet refuses to make that inference. Its policy line, straight out of the output JSON: "Visual changes are not interpreted as sounds or actions. Only editor-described cues enter the SFX list." Pixel markers go in a separate list called visual_markers_to_review, where they sit as questions rather than answers.

It generates nothing and spends nothing. It turns your decisions into a structured brief for the nodes downstream.

What it builds

Two things from the timeline: a continuous ambience bed and a list of timed effects.

You supply the ambience as free text - the default is quiet interior room tone, and yes, silence is a valid entry. The node will complain if you leave it blank, because "no ambience specified" and "silence" are different intentions and it won't guess which you meant.

Timed effects come from one of two places. If you fill in editor_cues_json with your own approved cues, those are used, sorted by time, each given an ID like SFX-01 and stamped editor-verified. If you don't, it falls back to events already marked editor_verified_event in the timeline - the ones that survived KIE • Review Video Events. Either way, only human-approved material becomes a cue.

dialogue_priority (default on) adds one instruction to the written brief: duck ambience and effects under speech.

Inputs and outputs

• events_json - forced input from KIE • Video Event Analysis (or the reviewed-events node). Needed for the duration and for the visual-marker list.
• ambience - multiline string, default quiet interior room tone. Capped at 1,500 characters.
• dialogue_priority - boolean, default true.
• editor_cues_json - optional, and the one you'll actually type in. The tooltip shows the format: [{"time_seconds": 4, "description": "door closes"}]. Up to 30 cues; each needs a time inside the video and a non-empty description.

Outputs:

• sound_cues_json - duration_seconds, ambience, dialogue_priority, the cues array, visual_markers_to_review (up to 30), and that policy note.
• sound_design_brief - the same thing as readable prose. This is what you paste into a Suno node, and what you read before you do.

Feed sound_cues_json into KIE • SFX Prompt from Cue to extract one effect or the ambience loop at a time, and hang sound_design_brief off KIE • Review Text so you actually look at it.

Where people get burned

"Describe the desired ambience; silence is also a valid description." Blank ambience field. Type silence if that's what you want - the node is asking you to be explicit, not to fill in a word.

"Provide at most 30 editor cues in a JSON array." You're past the cap.

"Editor cue 2 needs a time within the video and a description." A cue with a time past the video's duration, or a missing description. Times are seconds from the start.

Effect cues beyond the array. SFX Prompt from Cue indexes cues by number. If your sheet has three cues, cue_number: 4 errors rather than silently generating the wrong thing - which is the right behaviour, but it means re-running this node after adding cues is normal.

Unverified cues get rejected downstream. Only cues stamped editor-verified become SFX prompts. Your own editor_cues_json entries are automatically verified; markers from the pixel analysis are not, and never will be.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
cd ComfyUI-KIE-Nodes-Next
pip install -r requirements.txt

Restart, then look under KIE Next / Studio / Audio. This node is pure JSON shaping - no OpenCV, no API key, no credits. The audio-generation cost only starts when you connect its outputs to a Suno Sounds node, and the pack's template video-to-music-and-sound-brief.json is built to let you inspect every cue before you ever do.