TAG38

When “creating a song from text” is condensed into a single phrase, submitting finished lyrics and specifying the overall mood of a track can appear to be the same operation. When starting with a photo, it is likewise easy to confuse whether the image alone is sufficient or whether a short note is needed. This article is a step-by-step guide for short songs. Its purpose is not to speculate about the internals of the generation model, but to share the order in which inputs are read. WAV AI is assumed to be the hub for cross-navigation.WAV AI

Prerequisite: First classify what you already have

First determine which of the following you have on hand: (1) lyrics you have already written, (2) a single sentence describing a scene or mood, or (3) a photo, illustration, cover, or video still. If you press the generate button before finishing this classification, it will be difficult afterward to tell whether “the lyrics were overwritten or the image’s note was insufficient.”

Andrea Agostinelli and colleagues’ MusicLM: Generating Music From Text (Google Research, 2023) frames textual descriptions as control signals for music generation. It provides industry context for the ability to give direction in words, and is not a document explaining the workings of a specific product.

Step A: When you have no lyrics, or only want an instrumental (Inspiration)

Open Text to Music and use Inspiration

. This is the starting mode in which the entire track is described in a single prompt. It is suited to requesting background music for a video by describing the scene in one sentence, or to specifying the genre, mood, and sense of tempo for an instrumental without lyrics. If you write a prompt, limit it to information that actually changes the brief, such as genre, mood, instruments, sense of tempo, and the role within the video or song.

Step B: When the lyrics are already finalized (Custom)

In Custom on the same page, the lyrics field and music style field are separate. Put the lyrics in the lyrics field and the direction in the style field. The style field can also include elements such as genre, mood, instruments, sense of tempo, and role. It can be used to create either a song with lyrics or an instrumental without lyrics. This is where the use case of “adding music to lyrics” belongs. Avoid embedding finalized lyrics in a single-sentence Inspiration prompt, where they compete with the style specification.

Step C: When the image serves as the brief

If the image already has its own color and vitality, use Photo to Music.

1. Choose an image. It may be a photograph, illustration, cover image, or a still from a video.
2. Only when the image alone is musically ambiguous, optionally add music notes such as genre, tempo, instruments, and story. The notes are not required.
3. Create a song or an instrumental track.

Example uses include original music for video scenes (from the still), using artwork or a cover as inspiration for a song, and music for photo posts. You do not need to repeat in the notes what the image already conveys.

Jade Copet et al.'s Simple and Controllable Music Generation (MusicGen / Meta AI, NeurIPS 2023) argues that, in addition to being able to condition generation on text descriptions, generation can also be conditioned through a separate form of control. This provides the context for reading arrangements such as "separating lyrics and style" and "optionally placing notes beside the image" as "control design." This is not a claim that the product uses that model, nor does it treat the paper's numerical results as the product's performance.

Guidelines for deciding

• Lyrics are finalized → Custom. Put the lyrics and style in separate fields.
• Only a one-line scene, with no lyrics → Inspiration. Instrumentals can also be selected on the same page.
• The image carries the mood → Start with the image, and add notes only when it is ambiguous.

When finished, leave "inspiration / custom / photo" in the filename. That alone will let you explain the next day "which input type it was created from."