Take the running jokes, the trip disasters and the voice notes nobody answered, let a language model shape them into lyrics, and hear them sung over a full arrangement. Picasso IA runs both steps in the toolkit, as long as the people in the chat agreed to be in the song.
Every group chat has a plot: the running joke that never dies, the trip that went wrong, the friend who answers three days late. It already reads like lyrics, it just has no melody. A language model can shape the chat into verses, and a music model can sing them, so you hear your group chat as a full track. One condition comes first: everyone in the chat has to agree to be in the song.
A chat is a strange kind of diary. Nobody writes it to be kept, yet it holds the phrases a friend group actually repeats, the nicknames, the arguments about where to eat, the exact wording of a disaster. Those details are what make a song feel specific. A generic birthday song says you are great. A song built from your own chat says you once missed a ferry because of a sandwich, and everyone who was there will laugh at the second line.
The work splits cleanly into two jobs, and each one has models built for it. Picasso IA lists 502 models in the full catalog, among them 75 language models for drafting and rewriting text and 10 music generation models that turn lyrics plus a style description into audio. You run both in the toolkit, so the lyrics move from one step to the next without leaving the workspace. New accounts get free credits to try the whole chain once.
If you already have finished lyrics and only need the music step, the page on turning song lyrics into music covers that half in more depth.
Read this part before the rest. Do not paste other people's private messages into any tool without their consent. A group chat feels like your own material because you are in it, but half of the words belong to other people, and some of those words were typed on the assumption that they stayed in the chat. Anything you paste goes to the model you run it through, so treat the paste like forwarding the chat to a stranger.
The practical version takes two minutes. Ask the group, in the group, whether anyone minds. Offer to leave out any message they pick. Replace real names with nicknames or initials, and strip phone numbers, addresses, workplaces, health details, money talk and anything about someone who is not in the chat. If one person says no, that person's lines come out, or the project stops. A song is supposed to be a gift to the group, and one reluctant member turns it into something else.
The same rule covers the other direction. A song that mocks one person for the amusement of the others is not a tribute, however funny the chat found it at the time.
The most common reason these songs fall flat is that the lyrics try to retell the whole chat. A song has room for one idea and one hook. Decide what the song is about, then choose the musical style that suits it. The table below pairs common chat types with an angle and the style words that tend to fit.
| Chat type | Song angle | Style words for the music prompt | Watch out for |
|---|---|---|---|
| Running joke | Make the joke phrase the chorus | upbeat pop, singalong, playful, handclaps | The joke may need context a stranger lacks |
| Trip or event chaos | A story song with one verse per mishap | folk pop, road trip, acoustic guitar | Too many events crowd the chorus |
| Birthday or farewell | A tribute with one shared memory per verse | warm ballad, piano, heartfelt, gentle | Private details that do not belong in a song |
| Family chat | Gentle humor about routines and catchphrases | soft pop, nostalgic, light, bright | Sensitive topics sitting between the funny ones |
| Rivalry or roommate war | A mock-epic anthem about a tiny feud | cinematic, dramatic, tongue in cheek | Mockery that lands on one person |
Pick one row, not three. A chat can support several songs, so save the rest for a second one.
This is the full chain, from a chat everyone has agreed to share to a track you can play for the group. Budget an hour for the first song, mostly spent reading and editing lyrics rather than waiting on models.
The language model only knows what you paste, so its first draft will lean on whatever is most vivid in the excerpt. That is useful, and it is also the thing to steer. Tell it what the song is about in one sentence, name the line that should become the chorus, and say which details to leave out. A prompt like this works: write a four minute pop song about the weekend our friends missed the last ferry, the chorus is the phrase it was only one sandwich, keep the tone affectionate, and do not mention anyone by full name.
Then look at the draft the way the group would. A language model tends to smooth the odd phrases, and the odd phrases are the point. If your friend always writes the same misspelled word, put it back. If the model explained a joke instead of using it, ask for a rewrite that trusts the listener. Lyrics also need rhythm that a chat does not have, so expect the model to compress or reorder lines to make them singable.
Keep the lyrics under the limit and under control: music-2.6 takes up to 3500 characters of lyrics, but a song that tries to use all of them usually turns into a list. One angle, one chorus phrase that repeats, and a few verse lines per memory will sound better than the whole chat set to music.
For the music prompt, describe the sound the way you would to a friend with a guitar: genre, mood, tempo, instruments, whether the vocal should be male or female. You do not need to name an artist, and you should not try to copy a real singer. If you leave the lyrics empty, music-2.6 can write its own from the style prompt, but then the song contains nothing from your chat, which defeats the purpose here.
Some chats will not turn into good songs, and some should not become songs at all. These are the limits worth knowing before you spend credits.
In two steps, yes. A language model reads the excerpt you paste and writes lyrics from it, and a music model such as minimax/music-2.6 sets those lyrics to a melody with a sung vocal and instrumentation. What you get depends on how focused the lyrics are, so a chat excerpt with one clear angle works far better than a dump of hundreds of messages. Expect to edit the lyrics and regenerate at least once before it sounds right.
Only with their consent. Messages in a group chat are partly other people's words, and many were written on the assumption that they would stay in the chat. Ask the group before you paste anything, offer to remove lines a person picks, replace real names with nicknames, and drop private details. If someone says no, leave their lines out or skip the project. A song made as a surprise gift is still better made with permission than apologized for afterward.
This page uses minimax/music-2.6, a music generation model on Picasso IA that takes lyrics, a style prompt and structure tags such as [Verse], [Chorus] and [Bridge], and returns a finished track with vocals. The catalog also holds other music models, for example google/lyria-3 and elevenlabs/music, and each hears a prompt a little differently. Running the same lyrics through two of them is a cheap way to find the sound that suits your chat.
No, but you should still edit them. A language model such as claude-sonnet-5 or gpt-5 can draft verses and a chorus from the excerpt, and music-2.6 also has a lyrics optimizer that writes words from a style prompt when the lyrics field is empty. The first option uses your chat. The second does not, so the song would be generic. Your edit is what puts the specific phrases back in.
Less than you think. A few dozen messages around one event, one running joke or one person's moment is plenty, because the song has to fit a handful of verses and a chorus. If you paste months of conversation, the model produces a vague summary. Choose a stretch, tell the model the angle in one sentence, and save the rest of the chat for a second song if the first one works.
No, and it should not try. The vocal on a generated track is a synthetic voice shaped by your style description, such as male or female, warm or raspy, pop or folk. Imitating a real artist, or making a voice that sounds like a real friend without their agreement, raises problems that a funny group song does not need. Describe the genre and the mood instead, and you will get something original.
You can, but check two things first. Everyone named or quoted in the lyrics should be fine with a public version, which is a bigger ask than agreeing to a private joke. And each platform has its own rules about AI-generated audio and about what you can upload, so read the platform's current requirements before posting. For most groups, keeping the song in the chat where it started is the safer and better choice.
Language models handle many languages, so you can paste the chat as it is and ask for lyrics in the same language or another one. How well the sung result pronounces a given language, and how it handles slang or code-switching, varies by model and by prompt, so generate a short test before you commit to a final version. Spelling a hard word phonetically in the lyrics often fixes a pronunciation slip.
Do it before you paste, by hand. Copy the excerpt into a note, then swap every real name for a nickname or an initial, delete phone numbers, addresses, workplaces and anything about people who are not in the chat, and cut messages about health, money or relationships. Read the cleaned version once as if a stranger were reading it. A language model can also help you rewrite a line more vaguely, but the cleaning should happen first.
Each model run uses credits, and the cost depends on the model and the settings you pick, so any number written here would go out of date. Drafting lyrics with a language model is a small step, while generating the full track is the larger one, and regenerating adds to the total. New accounts start with free credits, and the pricing page is the place to read what the current plans include.
Your chat has been writing material for a song for a long time. Once the group says yes, open the Picasso IA toolkit and hear what it sounds like.
Choose your language
