Features Request - Audio quality score + Rename generated songs

[maysamassaf-beep]

Also if we can use ears to choose a quality generation, the possibility to score the audio quality can be of great help,
and another good addon would be the possibility to rename the generated songs
just suggestions btw , thanks for your work

[ServeurpersoCom]

I need to add a rename; but you can write the note or whatever you want in the renamed title (A rating function would be useful if this were a public site/music data) :)

[ServeurpersoCom]

I've added the renaming popup and a favorites heart that also serves as a mass delete button, now with a confirmation popup.
If simply adding your rating to the title isn't enough, it would be possible to add another icon to the right of the heart that allows you to rate the music. For example, from 1 to 5 (more granularity to know which song to rework), or a sad/neutral/smiling emoticon (but compared to the favorites, that only provides one more status). What do you think?

[maysamassaf-beep]

thanks for your effords and all the new options, btw i meant not a rating of the song, in comfyui i have a node that calculate the music quality with a score from 0 to 10 float analizing the song, so when you generate and see a 9.4 quality score it is because the audio was analyzed by that node that rated it very good and reported the score, not human rating... and you can be sure that the audio is of quality in this way despite our ears listening

[ServeurpersoCom]

Ah, that changes everything! I didn't understand that at all. You're talking about algorithmic quality analysis! Does the Python reference project allow for that? Not sure, but it's worth looking into!
For now I'm fixing a small WebUI bug for the favorites.

[ServeurpersoCom]

Response from a friend with whom we've been having fun testing algorithms from different papers for flow matching. We'll take a look at it :)

"Scragnog [TREK], — 15:22
I know the original python app includes a quality scoring mechanism, but when I tested it with HOT-Step 9000 it wasn't very useful, it could give a high rating and still result in a crappy generation.
I wonder how ComfyUI does it"

[maysamassaf-beep]

yes, i was convinced that the comfyui node i told about was directly related to ace-step and not just a common analyser because i noticed it go to analize the music , then the vocals, and anyway it runned very well into comfyui and for me was helpful thats why i suggested it, is not possible to look at that comfyui node code? to see if there is any difference with the original python app one... im a musician more than a coder so i can assure it was scoring the music very very appropriate

[maysamassaf-beep]

I taked a look, maybe can help, the node is named AudioQualityEvaluator and it is part of the pack JK-AceStep-Nodes

[scragnog]

@maysamassaf-beep is this the pack you're referring to - https://github.com/jeankassio/JK-AceStep-Nodes
I can't find anything to do with AudioQualityEvaluator in there

[maysamassaf-beep]

yes it is the node i was referring, and i repeat, im not a coder i just used it into comfyui and i was happy and the results was not wrong about the quality, so i suggested to add that as an option there but if for you it is unusefull no problem, patience, thanks, and sorry for have let you lost precious time, you was wondering how comfyui did it and i thought that a comparison with that module and the original python code you mentioned was a good idea...

[cynodesmus]

@maysamassaf-beep is this the pack you're referring to - https://github.com/jeankassio/JK-AceStep-Nodes I can't find anything to do with AudioQualityEvaluator in there

I believe we are talking about the code in ace_step_ksampler.py
#=======================================================================================================================

Audio Quality Evaluator Node

#=======================================================================================================================
class AudioQualityEvaluator:
"""
Evaluates decoded audio quality using LIBROSA (industry standard).
Professional-grade metrics for accurate scheduler comparison.

Returns a quality score from 0.0 (poor) to 1.0 (excellent).

Metrics (with librosa):
- Spectral Centroid Stability (timbre consistency) - 25%
- Spectral Rolloff (brightness/harshness) - 20%
- RMS Energy Variation (dynamics) - 20%
- Zero Crossing Rate (noisiness) - 15%
- Spectral Bandwidth (richness) - 10%
- Clipping detection - 10%

Librosa Features:
- Spectral centroid: Detects monotonic/robotic voice (CV < 0.1 = bad)
- Spectral rolloff: Detects muffled (<2000 Hz) or harsh (>8000 Hz)
- RMS variation: Detects flat/robotic (CV < 0.2) or glitchy (CV > 1.2)
- Spectral bandwidth: Detects thin/metallic sound (<500 Hz)

Fallback: Basic numpy metrics if librosa not installed
"""