How do you make a Spotify Canvas loop for your track with AI?

Spotify Canvas, as Spotify's own guidelines describe it at the time of writing
  length        3 to 8 seconds, looping
  shape         vertical, 9:16
  height        between 720 and 1080 px
  file          MP4 or JPG
  where it runs replaces the cover in the Now Playing view of the mobile app
  upload        Spotify for Artists > Music > your track > Add Canvas

That block is the whole job. To make a Spotify Canvas with AI, you take one image from your song, turn it into a still, animate the still in an image-to-video tool, trim the result to a few seconds, make the ending meet the beginning, and export a vertical MP4 no taller than 1080 pixels. Then you upload it per track in Spotify for Artists. At the time of writing, that upload is open to the first main artist on the track, or to a main artist on both the track and the release, with Admin or Editor access, and Spotify lets you add a Canvas to upcoming releases as well as live ones. Check the numbers above against Spotify's Canvas guidelines before you export, because platform specs move and this page will not move with them.

The AI part is the easy part. What decides whether the loop is worth looking at is the image you feed in, and that image should come out of the song, not out of the genre. A lo-fi track gets rain on a window by default because that is what a hundred other lo-fi tracks got. A lo-fi track whose chorus is about a porch light left on for two winters gets a porch light, and nobody else has that one.

Where this page sits among the others

A Canvas is not a music video. It is eight seconds at most, muted in the sense that it carries no sound of its own, and it plays over the Now Playing screen while the track runs. If you want a full clip for YouTube, with scenes and cuts across three minutes, read turning a track into a music video instead: different length, different tools, different problems. If you are still at the stage of getting the track onto Spotify at all, the release sequence comes first, because the Canvas upload only appears once the track exists in Spotify for Artists.

This page is for the narrow moment in between: the track is delivered or scheduled, you have the cover, and you want the thing that moves.

The image comes from the lyric, so get the lyric first

Here is the step most Canvas advice skips. Every guide tells you to "match the vibe of the track". A vibe is not something an image-to-video tool can render. An object is.

Take a chorus like this one:

[Chorus]
I left the porch light on for you
Two winters now, it's burning through
The moths all know your name by now
I left the porch light on somehow

The first and last lines are the same line with a different ending, and that repeated line names one physical object. That object is your Canvas. You do not need to explain the song, and Spotify's own guidance says to choose footage without talking, singing or rapping, because a looped clip will not stay in sync with the vocal. So there is no performer, no lip movement, no story arc. One porch light, some snow, two moths. Seven seconds.

If your song does not have a line like that yet, that is the gap we fill. The lyrics generator takes a short brief and returns a title and a full lyric with [Verse] and [Chorus] tags. Set the hook field to title-line and the title becomes the line the chorus is built on, which is usually the line that names the thing you will animate. One generation costs 20 credits, and at 100 credits to the dollar that works out to $0.20 a lyric inside a plan. There is no free trial: the cheapest way in is $5 a month, which brings 500 credits (25 lyrics) every month. If you release more than a couple of tracks, the lifetime plan is $19 once and comes with 1,500 credits a single time, enough for 75 lyrics. The generator does not sing and does not make video: it writes the words, and the audio and the loop happen in tools you choose.

Write the video prompt from the object, not from the mood

Once you have the object, the prompt almost writes itself. Most image-to-video tools accept a starting image plus a short text description of the motion. The mistake is to describe the feeling ("nostalgic, warm, longing"). The tool cannot animate longing. It can animate snow.

Image prompt (for the still):
  Vertical 9:16. A single porch light glowing above a weathered wooden
  door at night. Light snow in the air. Two moths near the bulb.
  Subject centred, plain dark background, nothing near the edges.
  No people, no text, no signs.

Motion prompt (for the image-to-video tool):
  Static camera. Snow falls slowly and evenly. The moths circle the
  bulb. The light flickers once, softly. No camera movement, no zoom,
  no new objects entering the frame.

Three details in that prompt do most of the work, and each one exists because of how Canvas is shown.

Static camera. A camera move has a start and an end, and a looped clip exposes both on every pass: the frame drifts left for six seconds and then snaps back. Motion inside a fixed frame (snow, smoke, water, a flag, a candle) loops far more forgivingly than motion of the frame itself.

Subject centred, nothing near the edges. Spotify's guidelines warn that some phones cut off the edges, and the Now Playing view also lays the track title and controls over the lower part of the picture. Anything important in the outer margin or the bottom third is a gamble.

No new objects entering. AI video tools like to add things. A cat walks in, a second light appears, the door opens. In a three-minute video that is a scene change. In an eight-second loop it is a glitch that comes back every few seconds for as long as the song plays.

Choose the loop type before you generate

Spotify's guidelines name three kinds of loop, and with AI footage the choice is practical, not aesthetic.

Continuous loop. The last frame flows into the first with no visible seam. This is the hardest one to get from a generator, because a generated clip rarely ends where it started. If your tool lets you set both a start frame and an end frame, give it the same still for both and you have a fair chance. If it does not, expect to fight the seam.

Hard cut. There is a clear edit point and you make it look intentional: a flash of the light, a gust of snow, a blink to black. This works when the motion builds toward something, so the cut reads as a beat rather than a mistake. Keep it gentle, because the same guidelines ask you to avoid rapid cuts and intense flashing.

Rebound. The clip plays forward, then backwards, then forward again. This is the one to reach for with AI footage, because it guarantees a clean seam: the last frame of the forward pass is the first frame of the reverse pass, by definition. It fails only when the motion has an obvious direction. Falling snow that rises again looks wrong. Smoke, drifting light, swaying grass, slow water and circling moths mostly survive it.

For the porch light, rebound is the right pick if the snow is sparse and slow, and hard cut is the fallback if the snow reads as clearly falling.

Trim, crop and loop it yourself

Whatever the tool exports, you will usually need to cut it down and force it into the Canvas shape. Generated clips drift: the opening seconds usually hold together, and somewhere after that the moths grow a third wing or the door changes colour. Take the cleanest few seconds and throw away the rest. The free command-line tool ffmpeg does all of it in two commands:

#1 take 3.5 clean seconds from the start, crop to 1080x1920, drop audio
ffmpeg -i generated.mp4 -ss 0 -t 3.5 \
  -vf "scale=1080:1920:force_original_aspect_ratio=increase,crop=1080:1920" \
  -an -c:v libx264 -pix_fmt yuv420p part.mp4

#2 rebound: play it forward, then reversed, for a 7 second loop
ffmpeg -i part.mp4 \
  -filter_complex "[0:v]split[a][b];[b]reverse[r];[a][r]concat=n=2:v=1:a=0[out]" \
  -map "[out]" -c:v libx264 -pix_fmt yuv420p canvas.mp4

Two passes of 3.5 seconds give 7 seconds, inside the 3 to 8 second window with room to spare. The crop line matters more than it looks: if the generator gave you a horizontal clip, it scales the clip up until it fills the vertical frame and cuts the sides off, which is exactly why you asked for the subject in the centre. If you started with a vertical clip, it mostly just resizes, trimming only whatever falls outside 9:16. yuv420p is there because some players refuse other pixel formats, and there is no reason to find out which ones.

If you prefer a video editor over a terminal, the same three moves apply in any of them: trim to the clean seconds, set a 9:16 project at 1080 by 1920, duplicate the clip and reverse the copy.

Watch the result on a phone, several times in a row, before you upload. A seam you do not notice once becomes the only thing you see on the fifth pass.

What the rules rule out

Spotify publishes a separate Canvas content policy, and at the time of writing it covers ground a generator will happily violate if you let it:

• no URLs, ticket vendors, social handles or other brands, and no Spotify logo
• no promotion of other songs, products, deals, contests or fan experiences
• text is allowed only as a secondary element; the image has to be the point
• explicit material (profanity, violence, drug references, sexual content, sensitive topics) has to be marked as such
• the general guidelines add: skip the song title and artist name, because the app already shows them

Image generators love to put lettering on signs, shop fronts and posters. That is why the image prompt above says "no text, no signs". If a word sneaks onto the door anyway, regenerate rather than hoping nobody reads it.

The rights question, briefly

Everything in the loop has to be something you are allowed to use. For AI footage that usually means two checks: the image or video tool's own terms on commercial use for the plan you are on, and whether anything in the frame belongs to someone else, like a recognisable person, a logo, a character or a photograph you fed in as the starting image. Tools differ on commercial terms and change them, so read the current page for the plan you actually pay for. This is general information, not legal advice; if the release matters commercially, ask someone qualified where you live.

The lyric side works the same way. Our own position is written down on the disclaimer: commercial use of generator output is allowed, we claim no authorship over it, and the rights are non-exclusive, so two users can end up with overlapping lines.

Five ways a generated Canvas goes wrong

It was generated horizontal. The crop keeps the middle third, and the middle third of a wide shot is often empty wall. Ask for vertical from the start, or at least for a centred subject.

It is too long. Eight seconds is the ceiling, not the target. Generators hold their shape for a few seconds and then start inventing. A tight three to four seconds, rebounded, looks better than eight seconds of slow melting.

It shows a face mouthing words. The clip will not stay in sync with the vocal, which is why Spotify advises against singing or talking footage. A face that is not speaking is fine. A mouth that moves will look like bad dubbing.

It does not match the cover. The Canvas replaces the cover art on the Now Playing screen, and Spotify's guidance suggests keeping it consistent with the cover and your artist imagery. If the cover is a flat red illustration and the Canvas is photoreal snow, the switch feels like a different release. Make the still from the same palette, or animate the cover itself. Designing the cover for the thumbnail helps here too: a cover with one big shape is a cover you can animate.

It is busy. A phone screen, a title overlay, playback controls and a moving picture, all at once. One subject and one kind of motion is plenty.

Where the Canvas fits in the release

In order: lyric, audio, master, cover, distribution, then the Canvas once the track shows up in Spotify for Artists. After that come pitching, pre-saves and watching the first numbers; what to do in Spotify for Artists after release covers that part. The Canvas is one of the few things you can change at any time without touching the release itself, so a weak first version costs you nothing but a replacement.

The one thing you cannot swap afterwards is the song. The loop can only be as specific as the line it came from, and a track whose words say nothing in particular leaves you animating rain on a window like everybody else.

So start from the words. If the track does not have them yet, or the chorus has no single object in it that you could film, write the brief, set the hook to title-line, and let the lyrics generator give you a title and a chorus to build on. Inside a plan each lyric costs $0.20; the way in is $5 a month, or $19 once for lifetime access with 75 lyrics included if you are putting out more than a single.