Best Apps to Animate a Photo in 2026: 7 Tools Compared

"Animate a photo" can mean three different jobs, and the app that is best for one is often useless for the other two. This page compares seven tools on what they actually animate, what their free plans give and where the watermark sits, read off each vendor's own pages on 13 and 15 September 2026. Then it uses 179 real talking photo renders on Percify to show what the job costs and why about 9 percent of photos fail whatever tool you use.

We make one of these tools, so every claim about the others comes from their own pages, linked and dated, and every number about ours from our own records.

Only one of the seven has no watermark on any plan, and only five of them make a face speak at all.

• Make the person in a photo talk, in your own voice, for public use: Percify, with no watermark on any plan.
• A few short talking photo tests: HeyGen's free plan, 3 videos a month of up to a minute.
• A talking or singing character from one image and your own audio: Hedra's Character 3.
• A talking photo trial for yourself: D-ID's 14 day trial, for personal use.
• Free talking photos where a watermark is fine: Vidnoz.
• Make an old family photo smile and blink: MyHeritage Deep Nostalgia.
• Make the whole scene move rather than a face talk: Runway, or an image to video model in Percify's model catalogue.

Getting this wrong is the most common reason people are disappointed, so it is worth being precise before comparing anything. Which of the three you want also decides the cost, and the credit pricing separates them clearly.

Motion on a still. The camera drifts across the photo, or, with an image to video model, the scene itself moves: hair, water, a slow turn. There is no speech. Runway's Gen-4 models do this, and so do the image to video models in Percify's model catalogue.

Face animation. The face moves: a blink, a smile, a small turn of the head, taken from a preset. There is no speech, so there is nothing to synchronise. MyHeritage's Deep Nostalgia does this for old family photos.

Talking photo. The person speaks your audio, with the mouth matched to it. This is the hard one, the expensive one, and the one people usually mean when they say animate my picture. Percify, HeyGen, D-ID, Hedra and Vidnoz all do this one, and most of this page is about it.

If you only want the first two, almost anything works and you should optimise for price. If you want the third, the differences between tools are real, and the lip sync comparison goes into which approach suits which source.

Across those 179 renders, grouped by the setting they ran at: The same records back the model catalogue, so any number here traces to the engine that produced it.

Read the middle column as a multiplier on your clip length. A 30 second talking photo finishes in about 3 minutes 49 seconds on the fast setting and about 12 minutes 34 seconds at 720p, for 2.4 times the credits. A later sample of 88 clips, June to September 2026, rendered faster on the fast engine, about 5.2 seconds per second of audio, so read these as the upper end.

That trade is the single biggest decision in this job, and it is not about quality in the way people assume. Higher resolution does not improve the synchronisation at all. It only puts more pixels on the mouth, which matters if the face is small in the frame and is wasted if it is not. You can check what those credits cost before choosing a setting.

Of the 160 unique source photos in that dataset: 145 were used once, 13 were used twice, one was used three times, and one was used five times. The same pattern shows up when replicating a video, where 86 percent of failures happen before generation starts.

Nobody runs the same photo five times because of a model problem. They do it because that particular photo cannot produce a good result, and nothing told them why. About 9 percent of source images get retried, and almost every retry is a source problem being mistaken for a tool problem.

This is the most useful thing in the data and the part no comparison article can give you, because it requires having run the job at volume. Before you judge any tool, check the photo.

Five properties decide it, and you can check all of them in under a minute. If your photo fails on face size specifically, a tighter crop is cheaper than paying for 720p.

1. Face size. The face should be at least a quarter of the frame width. Below that, the model has too few pixels of mouth to reconstruct, and this is when 720p stops being optional.
2. Nothing crossing the mouth. A hand, a microphone, hair, a scarf. The model cannot reconstruct what it cannot see, and this is a hard failure rather than a quality reduction.
3. Facing roughly forward. Past about three quarters profile, mouth geometry stops being reconstructable.
4. Even lighting on the face. Hard shadows across the mouth produce the same problem as something covering it.
5. One face only. These models expect a single subject. A second face gives unpredictable results, usually the wrong mouth moving.

A photo failing two of these is not worth retrying. Replace it. This is the same diagnosis path as the lip sync troubleshooting guide, because it is the same underlying model behaviour.

The mouth shapes come from phonemes in your audio. Unclear audio means ambiguous phonemes, and the mouth guesses. Most complaints that sound like video problems start here. Translated audio makes this harder still, which is why dubbing has its own timing rules.

Clean speech with no music under it, a consistent level, and real pauses between sentences. Normalising the track with ffmpeg ↗ costs nothing and removes a whole category of failure. If you are generating the voice as well, some voice tools preserve phoneme detail better than others for exactly this reason.

On the day we checked, only Percify had no watermark on any plan. HeyGen's free plan makes 3 videos a month of up to a minute. D-ID's trial is 14 days and 3 minutes, with a full screen watermark and a personal use licence. Vidnoz gives about 60 seconds a day at 720p with its watermark. Runway gives 125 one time credits, and Hedra and MyHeritage are free to start. That decides whether a result is usable more often than quality does, so check it before you invest time in a tool. Rights follow the output rather than the tool, so voice cloning and copyright is worth reading if the voice is not yours.

If you are only testing whether your photo works at all, watermarked output is fine, and several free options are good enough for that. If the video is going anywhere public, confirm the licence first, and what you may publish covers consent and disclosure, which are separate questions from licensing.

Singing. Sustained vowels and held notes sit outside what these models were trained on. Results are poor across every engine we have run.

Long clips from one photo. A still has one pose. Past roughly a minute, the absence of real body movement becomes obvious no matter how good the mouth is.

Strong accents against a mismatched language track. Phoneme sets differ by language and the model resolves toward what it knows, which is also why dubbing is its own pipeline rather than a setting.

Profile shots. Covered above, but it is the most common source failure so it is worth repeating.

Pick a photo that passes the five checks. Record or generate clean audio with real pauses. Run it on the fast setting first, because it costs a third of the time and 720p will not fix anything the source got wrong. Look at the mouth specifically rather than the overall impression. If it is wrong, change the photo before changing any setting, and only move to 720p when the face is genuinely small in frame. Where a limit is an engine limit rather than a product limit, no competitor claiming otherwise should be believed without a test, and the lip sync comparison covers which ones state theirs. If the result is going on social, the platform publishing rules decide how it must be labelled.

The whole loop costs about 66 credits on the fast setting for a half minute clip. You can try it ↗ before deciding, and the model list records which engine produced any given result so you can attribute it honestly. If you would rather work from a format that already performs, the flows are pinned pipelines built around specific shortform structures, and remaking a video shot for shot is the version of this job that starts from a reference rather than a photo.