Seedance 2.0 Mini
References input
AI fruit video model comparison
Which model gets
your talking fruit
story right?
We tested five video models on the same talking fruit scene. Watch the original clips, compare the acting and dialogue, and see what needs fixing before you publish.
The brief: a banana steps on a peel it hasn’t noticed, falls, then saves face with “Totally planned.” A skeptical strawberry replies, “Sure it was.” Can each model keep the joke, the characters and the subtitles intact?
See the shared prompt and character imageTry our pick: Seedance 2.0 Mini
Open the tested setup in Studio. Review or edit it before generating.
Same story, five models
Compare the fruit stories for yourself.
Watch with sound to compare the action, expressions and spoken joke. Each clip keeps the model’s original picture and audio.
MiniMax H3 Max
First frame input
Grok Imagine Video
First frame input
Veo 3.1 Lite
First frame input
Wan 2.7
References input
One take per model. Input modes and output settings vary; see how we tested.
Model-by-model review
What works, and what would you need to fix?
Each review covers the same twelve dimensions, from following the story to delivering a usable vertical clip. Use a model’s setup to try the same scene or adapt it to your own story.
Create with this setup
Opens this model with the test prompt, shared image and recorded settings. Edit them in Studio before generating.
Review 01References input
Seedance 2.0 Mini
ByteDance
Our first choice for the complete fruit scene.
The banana’s confidence turns into a believable accident, then an embarrassed attempt to save face. The strawberry’s reaction completes the joke, with separate speaking shots and readable subtitles.
- What worked
- The banana looks toward the strawberry before the accident. Its surprise, forced grin and thumbs-up sell the cover-up; the strawberry answers with a skeptical, amused expression. The cast stays consistent across all four shots, with readable, well-timed subtitles.
- What needs work
- An extra peel-like shape appears near the banana after the fall. The faces and outfits remain consistent, but the floor prop changes across the action.
- Dialogue & voice
- “Totally planned” accompanies the banana’s seated recovery; “Sure it was” accompanies the strawberry’s reply. The creator’s listening feedback was very positive about the voice performance, making this the strongest starting point for the complete spoken joke in this set.
Story & performance
- ✓Story adherence
- ✓Natural movement
- ✓Expressive acting
- ✓Shot sequence
Characters & world
- ✓Reference likeness
- ✓Character consistency
- ~Scene & prop continuity
Dialogue & voice
- ✓Dialogue accuracy
- ✓Dialogue timing
- ✓Voice & delivery
Subtitles & format
- ✓Subtitle accuracy & timing
- ✓Portrait framing
Before you publishThis is our first setup to try for this scene. Check the floor after the fall: the extra peel-like shape is the main continuity issue to resolve.
Compare this model across all twelve dimensions ↓Create with this setup
Opens this model with the test prompt, shared image and recorded settings. Edit them in Studio before generating.
Review 02First frame input
MiniMax H3 Max
fal Research · based on MiniMax H3
A strong alternative, with a less convincing slip.
The seated thumbs-up, strawberry reaction and matching subtitles make the exchange easy to follow. The fall is less convincing because the shoe does not clearly slip on the peel.
- What worked
- The proud opening, seated thumbs-up and amused strawberry reply carry the intended joke. All four shots arrive in order, the faces and outfits stay consistent, and both subtitle lines match the dialogue.
- What needs work
- In the wide shot, the foot moves near or over the peel before the banana lifts into the fall. The peel does not clearly slide under the shoe as requested.
- Dialogue & voice
- Both requested lines are present and arrive with the banana’s recovery and strawberry’s reply. The dialogue works in the story sequence; voice naturalness and character fit still await a listening review.
Story & performance
- ✓Story adherence
- ~Natural movement
- ✓Expressive acting
- ✓Shot sequence
Characters & world
- ✓Reference likeness
- ✓Character consistency
- ✓Scene & prop continuity
Dialogue & voice
- ✓Dialogue accuracy
- ✓Dialogue timing
- ?Voice & deliveryListening pending
Subtitles & format
- ✓Subtitle accuracy & timing
- ✓Portrait framing
Before you publishCheck whether the foot-to-peel contact is convincing enough for your joke. The reaction shots and captions work, but the cause of the accident is less clear.
Compare this model across all twelve dimensions ↓Create with this setup
Opens this model with the test prompt, shared image and recorded settings. Edit them in Studio before generating.
Review 03First frame input
Grok Imagine Video
xAI
Expressive slapstick that misses part of the story instruction.
The exaggerated backward tumble gives the scene a big physical reaction. But looking down at the peel before stepping changes the joke: the banana appears aware of the hazard instead of accidentally losing its confidence.
- What worked
- The broad backward tumble is readable as cartoon slapstick. A proud recovery and an amused strawberry close-up finish the scene. Character designs and outfits hold together across the four shots.
- What needs work
- “Sure it was.” appears during the banana close-up, before the strawberry speaking shot. The first subtitle also lingers after the first spoken phrase, so the text handoff needs attention.
- Dialogue & voice
- Both English lines are present, but the banana starts speaking around the tumble and recovery, before the requested seated speaking shot. The words are right; their placement weakens the pause before the cover-up. Voice performance still awaits a listening review.
Story & performance
- ~Story adherence
- ✓Natural movement
- ✓Expressive acting
- ✓Shot sequence
Characters & world
- ✓Reference likeness
- ✓Character consistency
- ✓Scene & prop continuity
Dialogue & voice
- ✓Dialogue accuracy
- ~Dialogue timing
- ?Voice & deliveryListening pending
Subtitles & format
- ~Subtitle accuracy & timing
- ✓Portrait framing
Before you publishRetime the dialogue and captions for a cleaner exchange. Keeping the original accidental setup would require a new take: the banana needs to miss the peel, rather than look at it.
Compare this model across all twelve dimensions ↓Create with this setup
Opens this model with the test prompt, shared image and recorded settings. Edit them in Studio before generating.
Review 04First frame input
Veo 3.1 Lite
The cast looks right; the action and spoken performance need work.
The characters look familiar, but the banana seems to crouch into a seated pose. Early dialogue, an unnatural voice performance and incorrect subtitles interrupt the intended comedy.
- What worked
- The reference fruit heads, outfits and kitchen remain recognizable, and the picture fills the portrait frame. The faces animate clearly, although the broad smiles convey the embarrassed cover-up and sarcasm less well.
- What needs work
- The descent looks like a deliberate crouch, so the accidental slip never reads naturally. Only three main shots appear, and the generated subtitles do not match the requested lines. This run was also shorter than the story timing requested.
- Dialogue & voice
- The banana’s line starts during the opening pose, and the reply arrives before the strawberry’s close-up. The creator also found the strawberry’s voice unnatural and masculine-sounding. Voice gender was unspecified, so the voice verdict concerns the unnatural delivery.
Story & performance
- ~Story adherence
- ×Natural movement
- ~Expressive acting
- ×Shot sequence
Characters & world
- ✓Reference likeness
- ✓Character consistency
- ✓Scene & prop continuity
Dialogue & voice
- ✓Dialogue accuracy
- ×Dialogue timing
- ×Voice & delivery
Subtitles & format
- ×Subtitle accuracy & timing
- ✓Portrait framing
Before you publishTry again with story timing that fits the selected duration. Check the accidental fall, voice performance and exact subtitle text before keeping the result.
Compare this model across all twelve dimensions ↓Create with this setup
Opens this model with the test prompt, shared image and recorded settings. Edit them in Studio before generating.
Review 05References input
Wan 2.7
Alibaba
A readable joke that needs work before a vertical release.
The fall and embarrassed recovery carry the joke, but a landscape scene sits inside the vertical file. The black bands make the characters much smaller on a phone, and the requested subtitles are missing.
- What worked
- The loss of balance and landing read as an accident. The banana's forced grin and thumbs-up, followed by the strawberry's amused reaction, carry the ending. Faces and clothing stay recognizable through four shots.
- What needs work
- The opening skips the requested flexing pose. More visibly, a landscape picture sits between large embedded black bands despite the portrait setting, and the requested English subtitles are absent.
- Dialogue & voice
- Both English lines are present and broadly align with the seated banana and strawberry reaction. The spoken exchange follows the ending; voice naturalness and character fit still await a listening review.
Story & performance
- ~Story adherence
- ✓Natural movement
- ✓Expressive acting
- ✓Shot sequence
Characters & world
- ✓Reference likeness
- ✓Character consistency
- ✓Scene & prop continuity
Dialogue & voice
- ✓Dialogue accuracy
- ✓Dialogue timing
- ?Voice & deliveryListening pending
Subtitles & format
- ×Subtitle accuracy & timing
- ×Portrait framing
Before you publishResolve the picture’s embedded black bands and add captions before publishing a vertical version. Check the resulting crop so both characters remain readable on a phone.
Compare this model across all twelve dimensions ↓The creator's checklist
Compare what matters for a fruit short.
Find the trade-offs that affect your scene: story adherence, believable performance, a consistent cast, clear dialogue and publishable framing. Every mark matches the corresponding model review.
✓Meets the check~Partly×Does not meet it?Listening pending
Swipe sideways to compare all five models →
| What matters for your story | Seedance 2.0 Mini | MiniMax H3 Max | Grok Imagine Video | Veo 3.1 Lite | Wan 2.7 |
|---|---|---|---|---|---|
| Story & performance | |||||
| Story adherence | ✓ | ✓ | ~ | ~ | ~ |
| Natural movement | ✓ | ~ | ✓ | × | ✓ |
| Expressive acting | ✓ | ✓ | ✓ | ~ | ✓ |
| Shot sequence | ✓ | ✓ | ✓ | × | ✓ |
| Characters & world | |||||
| Reference likeness | ✓ | ✓ | ✓ | ✓ | ✓ |
| Character consistency | ✓ | ✓ | ✓ | ✓ | ✓ |
| Scene & prop continuity | ~ | ✓ | ✓ | ✓ | ✓ |
| Dialogue & voice | |||||
| Dialogue accuracy | ✓ | ✓ | ✓ | ✓ | ✓ |
| Dialogue timing | ✓ | ✓ | ~ | × | ✓ |
| Voice & delivery | ✓ | ? | ? | × | ? |
| Subtitles & format | |||||
| Subtitle accuracy & timing | ✓ | ✓ | ~ | × | × |
| Portrait framing | ✓ | ✓ | ✓ | ✓ | × |
| Image input used | References | First frame | First frame | First frame | References |
| Create with this model | Try this setup | Try this setup | Try this setup | Try this setup | Try this setup |
Read across a row to compare a requirement you care about. A missing subtitle may be an editing task; a changed story intention or unnatural fall may call for another generation. Voice verdicts use creator listening feedback on Seedance and Veo; the other three await listening review. Read the evaluation rationale and evidence.
Why these dimensions matter
What makes a talking fruit story work?
A talking fruit short has to coordinate story intent, physical action, a recognizable cast and spoken performance. This scene puts those demands into one exchange: an unnoticed peel, an accidental fall, an embarrassed cover-up and a sarcastic reply. Our twelve checks identify which parts survive in the finished clip and what a creator would need to fix before publishing.
- Story & performance
- Intent, movement and acting each carry part of the joke. The banana must fail to notice the peel, fall by accident, then pretend to be in control. In the Grok take, looking at the peel changes that intention even though the tumble is expressive. We judge physical cause and effect within the cartoon style, including the shoe contact, loss of balance and landing.
- Characters & world
- Matching the shared image establishes the cast; keeping that cast and its world coherent through camera cuts sustains the story. These are separate checks. The Seedance take keeps recognizable faces and outfits, while an extra peel-like shape after the fall reveals a prop-continuity issue.
- Dialogue & voice
- Correct words, dialogue placement and vocal performance can succeed or fail separately. All five outputs contain the requested English phrases. In the Veo take, those words arrive before their intended visual beats. Voice delivery requires listening feedback; the transcript alone cannot establish whether a line sounds natural or lands the joke.
- Subtitles & format
- A finished short needs readable captions at the right moment and a picture that uses the requested frame. The Wan take illustrates why we inspect the image itself: its portrait file contains a landscape scene with embedded black bands. Missing subtitles create a second, separate publishing task.
The framework draws on VBench’s separation of identity, motion and prompt alignment; VBench-2.0’s treatment of physical plausibility and control; and VABench’s distinction between speech quality and audiovisual timing. We adapt those distinctions into editorial checks for this scene; the marks are not scores from their evaluation pipelines.
The findings you can take from this test
All five takes preserve a recognizable fruit cast and contain both requested English phrases. The differences appear in story intention, physical execution, speech placement, subtitles and vertical framing. Seedance 2.0 Mini comes closest to the complete brief in this set. The comparison table identifies the specific strengths and repair work for each output.
Scope: one formal take per model, with a shared story and image but different input modes and output settings. These findings describe the five recorded workflows; they do not estimate model-wide success rates. AI Fruit Video provides the tool and publishes this review.
Test inputs & run records
The story and image behind every clip.
Each model received the same intended scene and shared image, using the image-input mode recorded in its review. We preserved the original output, including action mistakes, generated subtitles and black bands, so you can compare the result with the brief.
Shared image · open full size ↗The banana-peel story brief
- 01The banana flexes, unaware of the peel.
- 02A confident step becomes an accidental slip.
- 03Seated, thumbs up: “Totally planned.”
- 04The strawberry replies: “Sure it was.”
Four shots, three cuts, two spoken lines and matching English subtitles. Both fruit characters and the kitchen should stay consistent.
Try the story yourself
Make your own talking fruit scene.
Start with the tested Seedance setup, or bring a new idea to Studio. You can edit the prompt and choose your settings before generating.