Bad Output / Poor Generation Quality (Mega Thread)
Shane
I'm consolidating all the 'bad output' posts together. This will be a lot easier for the devs to look into and potentially pick up more useful patterns of complaints and issues with the output.
Reference note to everyone:
The existing Udio models do not change in their output. They have not been worked upon or changed. I hope this is understood but I recognise that many users will still argue or say otherwise, this however can be demonstrated very clearly with seed output consistency, so we won't be getting into this debate or going down this rabbit hole with anyone as it's a waste of time*
The things and issues that are brought up here are things that will be looked into and worked towards improving in future models, not the existing ones. Lets look towards the future and hopefully the output should always keep improving with the next upcoming model updates. Nothing is garaunteed though, for how much and in what way things will get improved or not. Given the nature of GenAI it can be rather difficult to completely predict or control things in a precise manner.
*I say this but it is not 100% correct as there was a minor seed change on the 31st October. Which breaks full seed consistency to before that time but should not have affected the quality of output in any noticeable way and there should also still be some degree of seed consistency too. This is the only noticeable change that has been made to existing models though for a long time.
Shane
Merged in a post:
Bad sound quality with a lot of artifacts
Mario Ek Aparicio
I'm generating tracks using the new mlfem and ultra quality mode, but the generated audio, including the stems sound like very low quality mp3s with a lot of audible artifacts. I tend to set clarity to max and quality to ultra.
Shane
Merged in a post:
Prompts and tags are still ignored, and even if you specify laid-back ambient prompts and tags on a piano with a calm bass, high-pitched music with synths that aren't strange or creepy pianos is generated, and a completely different genre is generated, so it just consumes credit and time. More than hundreds of credits have been spent before. Why did this start to be an issue once you signed up for a subscription?
0203yt
Shane
Merged in a post:
vocal weirdness
budokan2732
Vocal Weirdness. the vocal quality seems to have dropped. vocals are coming out tinny, metallic resonance, the old down in a can sound. what's up with that? and too many defaults to rap/spoken even when listed in style reduction
0203yt
MusicFX also has the same prompt “Create a balanced sleep music track with a deep, low-pitched piano as the main instrument. Keep it slow (40-50 BPM), simple, and distraction-free. Focus on soft, resonant tones for comfort and comfort Avoid vocals or sharp sounds. When generated with “Add Subtle Ambient Noise Like a Faint Breeze or Reverb.”, it was generated with a high probability. SUNO also had a high probability that “piano music without bass synths, vocals, or strange sound effects was generated. Compared to other companies, there's no place that ignores prompts and tags as much as Udio and does creepy conversions.
0203yt
When generated with the same tags and prompts in Suno, the success rate was high. What's a waste is that SUNO has an electronic sound. However, I think it's more normal than Udio, which keeps ignoring prompts and tags.
Micah Maloney
Over a hundred credits burned through when it kept generating a voice speaking the lyrics, despite context of an earlier part where they were sung, specifying 'singer' in part of the prompt, in part of the lyrics '[male singer]'. and putting Spoken Word, Rap, and Hip Hop in unwanted genres and styles to avoid, just to try to keep it from generating more spoken lyrics. It was random chance that it finally got past to workable vocals again. Before upgrading to a paid subscription for 30 days, this wasn't happening nearly as often. Why has this started to become a problem this month?
Shane
Merged in a post:
ONLY VOCALS IN INSTRUMENTAL
Katarzyna Bitowt
Compose an instrumental traditional Viennese folk song (Heurigenlied) in C major with I–IV–V–I progression for the verses and a I–vi–IV–V progression for the chorus. Avoid any human voices, including choirs or acapella elements. Use a 3/4 time signature at 110 BPM. Use accordion for melody, guitar for rhythm, and violins for emotional interludes. Mood is witty, humorous and ironic. Do not use vocals, do not use lyrics. Only insturments. The mood should be humorous, inspired by Viennese wine tavers, similar to songs like "Ich trag im Herzen drin". Structure: 4-bar accordion intro, 8-bar verse with violin lead, and a 16-bar chorus with all instruments together.
Shane
Merged in a post:
Completely ignore prompts and tag specifications
0203yt
Completely ignore prompts and tag specifications
Unintended mixing of genres
Response to user feedback is uncertain
There is a big gap between reputation on SNS and actual performance
It might be wise to consider other AI services and continue to provide feedback. The lack of basic control functions is a serious problem.
prompts
INSTRUMENT: SOLO CONCERT HARP ONLY
NO OTHER INSTRUMENTS ALLOVED
TEMPO: 45 BPM FIXED
CLASSICAL HARP SOLO PERFORMANCE ONLY
[STRICT MUSICAL PARAMETERS]
Instrument: Solo Concert Harp Only
Genre: Pure Classical Ambient
Tempo: 45-50 BPM fixed
Style: Minimalist Classical
Dynamics: pp-mp only
[EXPLICIT EXCLUSIONS]
NO ELECTRONIC ELEMENTS
NO ROCK ELEMENTS
No vocals
No drums
NO SYNTHESIZERS
No modern day
[PERFORMANCE SPECIFICS]
- Simple descending arpeggios
- Long note duration (3+ seconds)
- Middle to low register only
- Sparse, peaceful harmonies
- Natural harp resonance only
tags
harp, relaxation, calming, peaceful,
Problems that occur:
dark ambient
hip-hop
synthesizer sound
space rock
is generated, and the specified pure harp solo is ignored
Expected results:
Only pure classical harp solo performances are generated
This issue is considered a serious bug related to basic control of AI models. Please investigate and correct it as soon as possible.
Shane
Merged in a post:
Problems with beat and rhythm.
Guillermo Eduardo Charcahuana López
I see it often: When creating a song, I think you should prioritize the beat or rhythm over the content or (lyrics, instruments or vocals)
Shane
Merged in a post:
almost ignore prompts and tags
0203yt
Even if prompts and tags are accurately entered and generated many times, they will be reflected in completely different genres and prompts and tags that have not been entered. Even if simple harp music is prompted, it is dark, electronic, and only unrelated things such as jazz pips or male vocalists come out, and even if slow and consistent tempo, between 40 to 50 BPM is input, it is generated with an uptempo, dirty, and muddy sound. Frankly speaking, it was a waste of time and credit. Even after a day, no good music is generated at all, and it's just a matter of deleting data. After all, the subscription will be canceled this month, and I will consider other subscriptions. It was a generative AI that was far from popular on SNS. It's a tool you can't generate in the first place.
Load More
→