Siren Scope (Audio → Image) 🧜

Audio can't go into Image Compare - a picture of it can. Siren Scope renders AUDIO to an IMAGE, and every design decision in it exists to make two takes comparable rather than to look nice. It's the visual checkpoint between "I generated a song" and "which of these three takes is the keeper," and it also hands gen_extra_info plain numbers that often answer the question faster than any picture.

Why it's built the way it is

Three choices do the heavy lifting:

• The dB scale is absolute, never auto-fitted. db_floor (default −80) and db_ceiling (default 0 dBFS) are fixed, so a quiet take renders dark instead of being silently boosted to fill the frame. Auto-normalizing each image is exactly what makes two spectrograms meaningless side by side - the whole point is that equal brightness means equal loudness.
• The pixel grid is fixed. The clip always spans width_px, so column x is the same moment in every render of a same-length track, and an A/B flip in Image Compare doesn't jitter.
• The bar grid lands where Siren Section would cut. Give it the same bpm / beats_per_bar / grid_origin_sec you gave the text encoder, and the bright lines are bar boundaries, dim ones beats - so you can pick a retake window off the picture and type those seconds straight into the Section node.

mode picks the view: mel spectrogram (structure, drop-outs, a band-limited top end), linear spectrogram (hard low-passes and resampling artifacts), waveform (level, silence, clipping), or mel + waveform stacked on one time axis. channels can draw a mono mix, one side, or both stacked - the way to catch a stereo collapse or one side dropping out.

The difference mode

Wire a second take into audio_b and it switches to A−B mode: black where the two are identical, bright where they diverge (warm = A louder, cool = B louder). Two independent takes never agree bin-for-bin, so diff_detail averages over musically-sized tiles first - musical (~1/8 of the mel bands × 120 ms, what a listener notices) is the default; fine (raw bins) is honest but usually unreadable. This is the direct check that a Siren retake did what it says: the frozen part of the track should come out pure black and only the marked section should light up. diff_span_db (default 6) decides how many dB of difference reaches full colour - leave it wide and two similar takes look identical when they aren't.

Settings worth knowing

n_fft (2048) and n_mels (192) - the 192 note again: at n_fft 2048 there are only 1025 frequency bins, and 256 mel bands would leave two permanently dead rows. Raise n_fft to 4096 before asking for more bands. db_floor is your inspect dial: raise toward −60 to see only what's audibly present, drop toward −100 to inspect the noise floor. time_labels off gives a clean image to overlay or diff pixel-for-pixel. And the gen_extra_info output carries plain numbers - loudness, crest, brightness, band balance, stereo - into a Generation Info dump, which is often the clearer answer to "how do these two differ" than squinting at colours.

Outputs: image (into Image Compare, like any other image), report, gen_extra_info.

Installing it

Part of Kinburg/Kinburg-Nodes - ComfyUI Manager search "Kinburg-Nodes", or cd ComfyUI/custom_nodes && git clone https://github.com/Kinburg/Kinburg-Nodes, restart. No extra dependencies - it analyzes an AUDIO input in place. If your two takes look identical in A/B, check diff_span_db and diff_detail before re-rendering; and if a retake's frozen section glows in difference mode, something re-noised audio it shouldn't have - that's a Siren Section problem, not a scope problem.