Video vs Speed
Hover to preview · click to pin. Top-left is high Elo and fast.
Creative Arenaby
Sprites
The AI video model leaderboard, tested on real ad briefs
Which AI video generator is best depends on the ad. Public arena Elo for the top video and image models, director stacks, and the Sprites bakeoff: the same ad prompts run on six video models and scored brief by brief, 56 clips in all.
Updated October 2, 2026
Video · public arena
Video models for campaign work
Shortlist generators for ads, UGC, and motion. Top cluster keeps full logos; hover a dot for the rest.
Video
Elo
Click a model to pin details · hover to preview
Reference-driven · audio on
Video ads
Built from reference images, a timed shot list, and native audio. Seedance 2.5 vs 2.0 Fast.
Brand intro · kinetic type · UI motion
Motion design
Claude Opus 5.5 wins every direction. For motion graphics and UI animation, text models writing code beat every video generator on brand fidelity, type, and easing: six on-brand takes at 60fps. The video models below are the losers on this brief.
Prompt
Premium motion design brand intro. On a deep navy background, bold white geometric shapes slide, snap and morph, assembling into the wordmark 'SPRITES'. Flat 2D vector, electric orange and lilac accents, smooth easing.Winner · Claude Opus 5.5 · code-rendered
Logo and character · smooth + hand-drawn
Video models · same prompt · lost
Phone selfie · product in hand
UGC creator ad
Seedance 2.0 Fast at 8.7. Natural creator delivery with the product readable in frame: the UGC default from this bakeoff.
Prompt
Vertical smartphone selfie, UGC style. A friendly woman in her 20s in a bright bathroom holds a white serum bottle to camera, tilts it to show the label, taps it and smiles while talking. Handheld, natural daylight, not cinematic.Hero packshot · liquid splash
E-com product ad
Kling 3.0, at the lowest price. For a real SKU, text-to-video is only for concepting: use the product photo as the start frame.
Prompt
Luxury product ad: amber perfume bottle, gold cap, label 'NOIR', on a wet black stone plinth. Water splash erupts in slow motion, warm rim light, camera orbits 90°. High-end commercial.Fall · shatter · fluid
Physics & action
Kling 3.0. Consistent with its #1 arena rank: best cause-and-effect physics, cheapest per clip.
Prompt
A hand knocks a full glass of red wine off a marble counter. It falls, hits the tiled floor and shatters, wine splashing outward. Slow motion, realistic fluid and glass physics.Image · public arena
Text-to-image leaderboard
GPT Image 2.5 leads; Seedream for bilingual packs; FLUX.2 when it must stay on-prem.
Image
Elo
Click a model to pin details · hover to preview
Image vs Speed
Hover to preview · click to pin. Top-left is high Elo and fast.
Still ads · GPT Image 2.5
GPT Image 2.5 for still ads
Five briefs marketers actually ship: product hero, UGC cover, display banner, carousel card, video end card.
Meta / Amazon · square
Product hero
GPT Image 2.5 wins on logo fidelity and multi-object control. Seedream 5.0 Pro is the bilingual packaging backup.
PromptStudio hero of a matte-black wireless earbud case on seamless peach, soft top light, readable logo on the hinge, room for headline top-left.
- 1
GPT Image 2.5Logo sharp, lighting commercial, crop-safe.9.4 - 2
Seedream 5.0 ProBest when the pack has CJK + Latin type.9.0 - 3MAI-Image 2.6Clean retail light; slightly softer mark.8.7
- 4
FLUX.2 MaxOpen-weight path; watch fine serifs.8.3 - 5
Nano Banana 2Fast enough for A/B volume.8.1 - 6Ideogram 4.0Over-designs the type block.7.6
TikTok / Reels cover
UGC hook still
Nano Banana 2 for speed + face naturalism. GPT Image 2.5 when the face must match a reference sheet.
PromptHandheld selfie of a creator holding the product toward camera, bathroom mirror light, candid, 9:16 crop, face sharp, product label readable.
- 1Nano Banana 2Natural skin, fast iterates.9.2
- 2GPT Image 2.5Best reference-face lock.9.1
- 3MAI-Image 2.6Slightly polished for true UGC.8.5
- 4
Grok Imagine 2.0Punchy crops for thumbnails.8.2 - 5FLUX.2 ProVolume option; faces drift.7.8
- 6Midjourney V8.1Too painterly for UGC.7.1
320×50 → 300×250 · type-heavy
Display banner
Ideogram 4 and Recraft win type-first layouts. GPT Image 2.5 over-illustrates unless constrained.
PromptFlat design banner: brand peach→pink gradient, bold headline "Cut agency retainers", small product glyph right, lots of negative space for live text overlay.
- 1Ideogram 4.0Headline kerning actually usable.9.3
- 2Recraft V4.1Vector-clean; brand kit friendly.9.0
- 3Seedream 5.0 ProStrong bilingual banners.8.6
- 4GPT Image 2.5Needs hard layout constraints.8.0
- 5FLUX.2 MaxGood shapes, soft type.7.7
- 6Imagen 4 UltraToo photographic for flat banners.7.2
Meta carousel · 1:1
Carousel card
GPT Image 2.5 for UI chrome. FLUX.2 Max when you must keep screenshots in-house.
PromptStep 2 of 4 carousel: before/after split of a Meta Ads Manager screen, crisp UI, muted chrome, orange callout on the winning campaign row.
- 1GPT Image 2.5Readable UI, correct density.9.1
- 2FLUX.2 MaxSelf-host for sensitive dashboards.8.6
- 3MAI-Image 2.6Clean but invents menu items.8.4
- 4Seedream 5.0 ProGood when UI is bilingual.8.0
- 5Nano Banana 2Fast drafts; check labels.7.9
- 6Recraft V4.1Better for illustrated steps.7.5
1080×1350 · logo + CTA
Video end card
Recraft + Ideogram for brand-safe geometry. Midjourney loses the mark.
PromptEnd card: Sprites wordmark centered on #F8F8F8, peach glow behind, single CTA pill "Start free trial", no extra ornaments.
- 1Recraft V4.1Logo geometry holds.9.4
- 2Ideogram 4.0CTA type crisp.9.1
- 3GPT Image 2.5Needs negative prompts to stay flat.8.5
- 4Seedream 5.0 ProPalette-faithful.8.3
- 5FLUX.2 ProSlight mark drift.7.9
- 6Midjourney V8.1Beautifully wrong logo.6.8
Orchestrated · Opus 5.5 & Astra
Director stacks
Same base models, but Claude Opus 5.5 or GPT-6 Astra writes the shot list, picks references, and runs a revise pass.
Orchestrated
Stack EloEditorial
Click a model to pin details · hover to preview
Orchestrated vs Speed
Hover to preview · click to pin. Top-left is high Elo and fast.
Sprites bakeoff · round 1
Same brief. Six models. Our reel.
Identical prompts, scored /10, for when you care how a model behaves on marketing briefs. Click a bar for that model’s best clip.
Overall
PreferenceAvg / 10
Click a bar to open that model’s best bakeoff clip
Video vs Price
Hover to preview · click to pin. Top-left is high Elo and cheap.
Close-up portrait · micro-expression
Face realism
Seedance 2.0. Even in Fast mode it passed all six face checks. The full rubric is in How we rate a face.
Prompt
Extreme close-up portrait of a woman in her early 30s with freckles and natural skin, soft window light. She looks down, lifts her eyes to camera and breaks into a shy smile; subtle eye movement and a blink. 85mm, photorealistic.Orchestrated vs Price
Hover to preview · click to pin. Top-left is high Elo and cheap.
Image vs Price
Hover to preview · click to pin. Top-left is high Elo and cheap.
Trend template · performance to camera, seamless backdrop
Rapper on orange
Looks right, isn't the trend. Seedance gets the set, but the format lives on a real verse and delivery synced to audio. Round 2 drives it with a track and a performance reference.
Prompt
A charismatic young rapper in an oversized black hoodie and silver chain performs to a static camera against a seamless bright orange backdrop, head nods, expressive hand gestures, rapping with energy. Clean studio light.Trend remake · four methods
Recreating the orange-booth trend
The same duo routine made four ways, from a still start frame up to a full character swap.
Trend template · synchronised crowd, still centre
The “Storm” crowd dance
No usable take. The trend is the synchronised choreography around a frozen lead. None of the three delivered it.
Prompt
Wide symmetrical shot in a grey school gym: dozens of young men in navy uniform blazers do a rigid, perfectly synchronised routine, while one pale young man in a white shirt stands still and stoic in the centre. 16mm grain, arthouse music-video look.Code-rendered · ~$0.35–$1.80 a clip in Opus tokens
UI launch, three variations
Written and rendered in code at 1080×1350, pixel-exact to the brand.
Playbooks
Defaults for marketing workflows
What we reach for when the job is ads, UGC, or stills. Not a generic model ranking.
| Job | Default | Workflow |
|---|---|---|
| Motion design / logo intros | Claude Opus 5.5 | Write the motion in code (Remotion / CSS / Canvas). Video models lose on type, brand, and easing for motion graphics and UI. |
| Faces / portraits / talent | Seedance 2.0 (std) | Use image_references for identity across a campaign; Veo 3.1 when native dialogue is needed. |
| E-com product ads | Kling 3.0 | Product photo as start_image to lock the label; Marketing Studio for one-click product ads at scale. |
| UGC creator ads | Seedance 2.0 Fast | Seedance 2.0 for silent B-roll hooks; Seedance 2.5 when you need refs + longer takes with audio. |
| Physics / action / food | Kling 3.0 | Best cause-and-effect; cheapest. Seedance 2.0 for pure liquid splashes. |
| Viral clone (Storm, Hotel Lobby, etc.) | Genjutsu motion transfer | Driving clip + brand faces as image references. Not text-to-video. |
| Performance / music clips | Seedance 2.0 | Add an audio reference so gestures and lips follow the track. |
| Talking-head ads with VO | Opus → Seedance 2.5 | Opus shot list + reference stills; Seedance 2.5 for identity lock and audio. Astra → Omni Flash when you need a cheaper one-shot VO. |
| Still product heroes | GPT Image 2.5 | Seedream 5.0 Pro when the pack has CJK type; FLUX.2 Max if it must stay on-prem. |
| Feed volume stills | Nano Banana 2 | Fastest AA-tier still. MAI-Image 2.6 when you want cheaper near-frontier quality. |
Every bakeoff model, clip by clip
Each model page has every clip it made, the exact prompts, its score and place on each brief, and where it breaks on ads.
Kling 3.06.9/10 on 7 briefs, 2 won, $0.28 a clip.Seedance 2.0 Fast6.7/10 on 7 briefs, 3 won, $0.47 a clip.Veo 3.1 Fast6.9/10 on 4 briefs, 0 won, $0.90 a clip.Wan 3.06.7/10 on 5 briefs, 0 won, $0.33 a clip.MiniMax H3 Max7.3/10 on 4 briefs, 0 won, $0.47 a clip.Happy Horse7.1/10 on 7 briefs, 0 won, $0.47 a clip.
Head-to-heads and prompts
Kling 3.0 vs Veo 3.1 for adsKling 3.0 vs Veo 3.1 on the same ad prompts: the clips side by side, the score on each brief, which to use for product, UGC and faces, and the price.Seedance 2.0 vs Veo 3.1 for adsSeedance 2.0 vs Veo 3.1 on the same ad prompts: the clips side by side, the score on each brief, which wins faces and UGC, and when Veo is worth more.Kling vs SeedancePlans, cost per finished ad and failed-generation policy, side by side.AI video prompt examplesSeven AI video prompts for ads, copied word for word, each run on Kling, Seedance, Veo, Wan, MiniMax and Happy Horse, with every clip and its score.AI video tools for adsHiggsfield, Runway, Kling, Seedance and the avatar tools on cost per finished ad.
Frequently asked questions
Which AI video generator is best in 2026?
It depends on the job. On public preference, Gemini Omni Flash leads the Artificial Analysis text-to-video board with audio. On ad briefs, Seedance 2.0 wins faces and UGC creator ads, Kling 3.0 wins product packshots and physics, and Claude Opus 5.5 writing code beats every video model on logo intros and motion graphics.
Which AI video model is best for product ads?
Kling 3.0. On the luxury packshot brief it kept the label, the splash and the set, scoring 8.5/10 at $0.28 a clip, the lowest price in the round. For a real SKU, give it the product photo as the start frame. For a creator holding the product, Seedance 2.0 Fast won the UGC brief at 8.7/10.
Is there a leaderboard that ranks AI video generation models?
Yes. The Artificial Analysis Text-to-Video Arena and Arena.ai rank models on blind human votes. The Creative Arena puts that public Elo beside its own bakeoff, where every model gets the same ad prompt and every clip is scored, because a model that wins general preference can still lose an ad brief.
Sora 2 vs Veo 3: which should I use for ads?
Veo. OpenAI closed the Sora app in April 2026 and removed Sora 2 from its API on September 24, 2026. Veo 3.1 is the pick when the ad needs dialogue generated in the clip; for faces and UGC without dialogue, Seedance 2.0 scored higher in the bakeoff.
What is the best image model for ad creatives?
GPT Image 2.5 leads the Artificial Analysis text-to-image board and wins product heroes and carousel cards (9.4/10 on the product hero brief). Nano Banana 2 is the fast pick for UGC hook stills, and Ideogram 4.0 and Recraft V4.1 win type-heavy banners and end cards.
Which AI video generator makes the most realistic faces?
Seedance 2.0. Even in Fast mode it passed all six face checks and scored 9/10 on the close-up portrait brief, ahead of Veo 3.1 Fast at 8.5/10.
Which AI video model handles physics best?
Kling 3.0, at 9/10 on the falling wine glass: tip, fall, shatter and spread, in order, with believable shards. Happy Horse was second; Seedance floated the glass and Wan never broke it.
What is the best AI for motion graphics and logo intros?
A text model writing the animation in code. Claude Opus 5.5 won every direction of the motion brief. The video models lost on type and brand: one spelled the wordmark 'SPRITS'.
Should an LLM write the shot list before the video model runs?
For ads, yes. Every director stack on the board scores above its base model alone. The top stack is Opus → Seedance 2.5: Claude Opus 5.5 picks reference stills and shot timing, and Seedance 2.5 holds identity across a long take.
How are the Creative Arena rankings made?
Public Elo comes from the Artificial Analysis video and image arenas, snapshot 2026-09-24. The bakeoff gives every model the identical prompt per ad brief and scores each clip out of 10 for that brief. Director stack scores are editorial.
Want finished ads, not model bakeoffs?
Director by Sprites: every monday you get finished, on-brand video ads, reviewed by a human. You never write a prompt. You pick the winners. Sprites ships the campaigns.
Apply for Director