# Best image to video AI in 2026: one photo through 5 models
By TokLab · Published 2026-11-04 · Updated 2026-11-04
Image to video looks easy in a demo. You upload a photo, describe a movement and get a clip. With your own photo, the questions are more specific: does her face stay the same, does the movement you asked for actually happen, and does the model quietly change the frame?
We gave one photo to five image to video AI models on November 4, 2026 and asked each for the same three movements. All five clips together cost $0.8105. Each model's clip is below, next to a checklist of what it did.
TokLab, which we make, sells all five models, and the clips were rendered through its API. Seedance 2.5, which leads the public image to video leaderboard, isn't in the test; the reason is near the end.
## Quick answer: what is the best image to video AI?
Of these five, **Seedance 2.0 Mini** followed the prompt most completely. It was the only one to lower the umbrella. **MiniMax H3 Standard** came closest for about 40% of the price, at $0.0769 a clip, and kept the photo's exact shape. If your photo is 16:9 already, **Veo 3.1 Fast** gives you 8 seconds for $0.168. Pick it for other shapes and it adds black bars.
On the Artificial Analysis leaderboard, where viewers vote blind between clips, the top image to video model is Seedance 2.5.
## The photo and what we asked for
The start frame is a 3:2 photo of a young woman laughing under a clear umbrella on a rainy street at night. We made it in our [AI image generator test](https://toklab.ai/blog/best-ai-image-generator). Every model got it as the first frame, with this prompt:
"She lowers the clear umbrella slightly, keeps laughing and turns to look down the street as the rain keeps falling; a car passes behind her and its headlights sweep across the wet pavement; slow push-in"
That's three things you can check by watching: an object moves (the umbrella comes down), she moves (she turns toward the street), and something happens behind her (a car goes by). Each model ran once, at its lowest resolution that still shows her face clearly.
## Image to video AI models at a glance
| Model | Best for | Clip we got | Billed | Image to video Elo |
| ------------------------------------------------------ | ---------------------------------- | ------------- | --------: | -----------------: |
| [Seedance 2.0](https://toklab.ai/docs/seedance-2-0) Mini | following the whole prompt | 752×560, 5 s | $0.185558 | 1135 (2.0) |
| [MiniMax H3](https://toklab.ai/docs/minimax-h3) Standard | animating photos in bulk | 1152×768, 5 s | $0.076926 | 1130 (H3, 768p) |
| [Grok Imagine Video 1.5](https://toklab.ai/docs/grok-imagine-video-1-5) | keeping her face toward the camera | 672×448, 6 s | $0.24 | 1049 |
| [PixVerse v6](https://toklab.ai/docs/pixverse-v6) | quick drafts without sound | 1024×688, 5 s | $0.14 | 958 |
| [Veo 3.1](https://toklab.ai/docs/veo-3-1) Fast | photos that are already 16:9 | 1280×720, 8 s | $0.168 | 936 |
Source: Artificial Analysis Video Arena, image to video leaderboard, November 4, 2026, next to the amount TokLab charged us for each clip that day. A tier in brackets means the leaderboard rates a different version of the model from the one in our clip.
## Seedance 2.0 Mini: best image to video AI for following the prompt
| Check | Result |
| ------------------------ | --------------------------------------------------- |
| Face kept | yes; her freckles fade a little by the third second |
| Umbrella lowered | yes, the only model that did it |
| Turned toward the street | fully, until her face leaves the frame |
| Car passed | yes |
| Frame shape | 752×560, a little squarer than the photo |
| Sound | yes |
| Render time and cost | about 2 minutes, $0.185558 |
Seedance 2.0 Mini did the most of what we wrote. Around the middle of the clip the umbrella drops to her side, and she keeps turning until we're looking at the back of her head and a car's lights on the right of the frame.
That's the literal reading of "turns to look down the street", and it's also the catch. By the last second her face is gone. If you're animating a portrait for a profile video or an ad, you probably want her to glance at the street and come back, so say that in the prompt.
The face itself held up. The only change we noticed was that her freckles got fainter and her hair looked a touch shorter by the third second. The clip is also slightly squarer than the 3:2 photo, so expect a small crop if you place it back next to the original.
## MiniMax H3 Standard: best image to video AI for animating photos in bulk
| Check | Result |
| ------------------------ | ----------------------------- |
| Face kept | yes |
| Umbrella lowered | no, it stays overhead |
| Turned toward the street | fully |
| Car passed | yes, a silver car on the left |
| Frame shape | 1152×768, exactly 3:2 |
| Sound | yes |
| Render time and cost | about 1.5 minutes, $0.076926 |
MiniMax H3 Standard returned a 768p clip and kept the photo's 3:2 shape to the pixel, so nothing was cropped or added. She laughs, turns until she's facing down the street and her face leaves the frame, and a silver car rolls past behind her with its lights on the wet road.
It ignored the umbrella. It stays exactly where it was in the photo the whole time.
At $0.076926 a clip it costs about 40% of Seedance 2.0 Mini, which makes it the one to try first on a batch. If you're animating a hundred photos, keeping the original framing is worth more than one missed instruction. The leaderboard's MiniMax H3 entry, at 1130, is the 768p tier, close behind Seedance 2.0.
## Grok Imagine Video 1.5: best for keeping her face toward the camera
| Check | Result |
| ------------------------ | ----------------------------- |
| Face kept | yes |
| Umbrella lowered | no |
| Turned toward the street | partway, face stays on camera |
| Car passed | yes |
| Frame shape | 672×448, exactly 3:2 |
| Sound | yes |
| Render time and cost | 57 seconds, $0.24 |
Grok Imagine took the opposite approach to Seedance and MiniMax. The camera pushes in clearly, she turns her head toward the street but keeps her face to us, and a car slides by in the background. For a portrait, that's arguably the more useful result.
It also kept the photo's 3:2 shape, and it was the fastest of the five to finish, in 57 seconds. The trade-off is resolution and price: 672×448 for $0.24, the most expensive clip here, partly because Grok's shortest clip is 6 seconds.
## PixVerse v6: best for quick drafts without sound
| Check | Result |
| ------------------------ | -------------------------------------- |
| Face kept | yes |
| Umbrella lowered | no |
| Turned toward the street | turned away, then back to the camera |
| Car passed | yes |
| Frame shape | 1024×688, within a few pixels of 3:2 |
| Sound | no, audio is a paid option we left off |
| Render time and cost | about 64 seconds, $0.14 |
PixVerse v6 made the most of the turn as a gesture: she looks over her shoulder as a car passes, then turns back and laughs at the camera. It reads like a natural moment, though the prompt didn't ask for the return.
The clip is silent because PixVerse charges extra for audio and we didn't switch it on. Its shape is within a few pixels of the photo's, so nothing visible was cropped.
## Veo 3.1 Fast: best for photos that are already 16:9
| Check | Result |
| ------------------------ | --------------------------------------- |
| Face kept | yes |
| Umbrella lowered | no |
| Turned toward the street | partway |
| Car passed | yes, a dark car |
| Frame shape | 1280×720, with black bars on both sides |
| Sound | yes |
| Render time and cost | about 2 minutes, $0.168 |
Veo 3.1 Fast returned the longest clip, 8 seconds at 720p, and the motion is smooth: a partial turn of her head, a dark car passing, rain on the umbrella. It's a good clip, sitting inside black bars.
Veo placed our 3:2 photo in the middle of a 16:9 frame and filled the sides with black. You can crop them out, but then you're back to a smaller clip than you paid for. If your photos are already 16:9, this doesn't come up, and 8 seconds for $0.168 is good value. Every Veo 3.1 tier renders 8 seconds per clip.
## Are you comparing apps or the models inside them?
Most image to video apps run one of these models. ByteDance's Dreamina runs Seedance, Hailuo is MiniMax's app, Google's Flow runs Veo, and Grok Imagine lives inside Grok. When two apps give you very similar clips from the same photo, they're often using the same model.
That's why this test compares models instead of apps. The app decides the interface, the free credits and the watermark. The model decides whether the umbrella comes down.
## Where image to video still falls short
**Faces in group photos.** [Nemovideo's review](https://www.nemovideo.com/blog/best-image-to-video-ai) of image to video tools calls face stability "still a legitimate problem" and says group photos are the hardest case. With one face, all five models here kept her recognizably the same person. Test a group photo separately before you trust a model with one.
**Object instructions.** All five models turned her head and added the car. Four of them left the umbrella exactly where it was. Movements of the person seem to get through more reliably than movements of something she's holding.
**The shape of your photo.** Three models kept the 3:2 photo's shape, or within a few pixels of it, one made it slightly squarer, and one added black bars. If you're about to batch a hundred photos that aren't 16:9, run one first and check the edges.
## Image to video apps on other lists
Runway, Pika and Luma appear on most lists of image to video tools. They weren't part of this test; these details are from their pricing pages on November 4, 2026.
| App | Free option | Paid from |
| ------ | ---------------------------------------- | ---------------------------------- |
| Runway | 125 one-time credits | $15 a month, $12 billed yearly |
| Pika | free credit packs, no commercial license | $10 a month (Starter, 900 credits) |
| Luma | none listed | $30 a month (Plus, 10,000 credits) |
If you try one with the photo you need to animate, run the same three checks as above: did the face stay, did the action happen, did the frame keep its shape.
## The model we didn't run
Seedance 2.5 leads the Artificial Analysis image to video board at 1159, with Seedance 2.0 next at 1135. A 4-second Seedance 2.5 clip at 480p costs about $0.41 on TokLab, more than any clip in this test and about half of what all five cost together, so this round used Seedance 2.0 Mini instead. Kling 3.0 and Wan 3.0 aren't in it either. If you have budget for one premium render of your final shot, Seedance 2.5 is the one the leaderboard points to.
## FAQ
### Which image to video AI keeps the face the same?
All five models in this test kept a single face recognizable. Seedance 2.0 Mini softened her freckles slightly. Face changes get worse with group photos, so test those separately.
### How long can an image to video clip be?
The clips here ran 5 to 8 seconds. Most of these models go up to 15 seconds in one pass, and Seedance 2.5 up to 30.
### Is there a free image to video AI?
Yes, with limits. Dreamina advertises free image to video with Seedance, and most other free tiers cap length, resolution or the number of clips a day. Current TokLab rates for the models above are on the [pricing page](https://toklab.ai/pricing).