AI models
AI image and video models, compared.
You never have to choose one.
Misu routes each shot to the image or video model that fits it — then holds your product and your cast steady across all of them. Here is what each one is for.
Looking for fashion talent to cast, not the generation engines underneath? See the AI model agency.
01
Compare at a glance
Every model Misu routes to, with what it's actually built for and what it costs in tokens. Exact cost shows in the Studio before you run.
- Seedance 2.5Video · ByteDance
- Best for
- Long single takes. Any length from 4 to 30 seconds, with no cut to hide.
- Cost
- ~263 tok/sec (480p–720p)
- Seedance 2.0Video · ByteDance
- Best for
- Campaign film and fashion motion up to 15 seconds.
- Cost
- ~168 tok/sec (~135 Fast)
- Kling 3.0Video · Kuaishou
- Best for
- Product turntables and detail shots where the object has to behave like a real object.
- Cost
- ~94 tok/sec
- Veo 3Video · Google
- Best for
- Campaign film where atmosphere and lighting do the work.
- Cost
- ~223–417 tok/sec
- MiniMax H3Video · MiniMax (Hailuo)
- Best for
- 2K clips with native stereo audio, from 5 to 15 seconds.
- Cost
- ~145 tok/sec (~45 Max)
- ViduVideo · ShengShu
- Best for
- Putting several characters in the same frame.
- Cost
- ~223 tok/clip
- Nano Banana ProImage · Google
- Best for
- Instruction-led production work: precise, structured prompts carried out literally.
- Cost
- ~84 tok/image (2x at 4K)
- GPT ImageImage · OpenAI
- Best for
- Editorial and campaign stills with strong prompt adherence.
- Cost
- ~30 tok/image (Standard)
- SeedreamImage · ByteDance
- Best for
- Holding one character across a multi-shot campaign, on Seedream 4.5 and 5.0 Lite.
- Cost
- ~17–75 tok/image
- Flux.2Image · Black Forest Labs
- Best for
- Photoreal hero shots and campaign stills, on Flux.2 [pro].
- Cost
- ~18–23 tok/image
| Model | Maker | Type | Best for | Cost |
|---|---|---|---|---|
| Seedance 2.5 | ByteDance | Video | Long single takes. Any length from 4 to 30 seconds, with no cut to hide. | ~263 tok/sec (480p–720p) |
| Seedance 2.0 | ByteDance | Video | Campaign film and fashion motion up to 15 seconds. | ~168 tok/sec (~135 Fast) |
| Kling 3.0 | Kuaishou | Video | Product turntables and detail shots where the object has to behave like a real object. | ~94 tok/sec |
| Veo 3 | Video | Campaign film where atmosphere and lighting do the work. | ~223–417 tok/sec | |
| MiniMax H3 | MiniMax (Hailuo) | Video | 2K clips with native stereo audio, from 5 to 15 seconds. | ~145 tok/sec (~45 Max) |
| Vidu | ShengShu | Video | Putting several characters in the same frame. | ~223 tok/clip |
| Nano Banana Pro | Image | Instruction-led production work: precise, structured prompts carried out literally. | ~84 tok/image (2x at 4K) | |
| GPT Image | OpenAI | Image | Editorial and campaign stills with strong prompt adherence. | ~30 tok/image (Standard) |
| Seedream | ByteDance | Image | Holding one character across a multi-shot campaign, on Seedream 4.5 and 5.0 Lite. | ~17–75 tok/image |
| Flux.2 | Black Forest Labs | Image | Photoreal hero shots and campaign stills, on Flux.2 [pro]. | ~18–23 tok/image |
02
Video models
Image-to-video, reference-to-video and text-to-video — for product motion, campaign films and social cuts.
- ByteDanceSeedance 2.5ByteDance's longest single-shot model. Thirty seconds in one take, with the sound made in the same pass as the picture.
- ByteDanceSeedance 2.0Cinematic motion and sound for clips up to 15 seconds.
- KuaishouKling 3.0The product motion model. Physical realism for spins, detail shots and garments, in three tiers.
- GoogleVeo 3Google's cinematic video model. The pick when light, lens and mood carry the shot.
- MiniMax (Hailuo)MiniMax H3Hailuo 3.0. Clips of 5 to 15 seconds at 2K, with stereo sound on every generation.
- ShengShuViduSeveral subjects, one shot. Give Vidu up to seven images and it composes them into a single clip.
03
Image models
Stills: product photography, on-model imagery, typography and edits.
- GoogleNano Banana ProGoogle's instruction-following image model. Say exactly what to change, and it changes that.
- OpenAIGPT ImageOpenAI's image models, from everyday shots to campaign finals. Flare for speed, Sunburst for control.
- ByteDanceSeedreamByteDance's image family. One model for dense, text-heavy heroes, two for keeping a face across a campaign.
- Black Forest LabsFlux.2Photorealism and close prompt adherence from Black Forest Labs. Kontext Max edits what you already have.
04
Questions people ask
Which AI model should I use for a product turntable or detail shot?
Kling 3.0. Its Pro and Omni tiers are built for product motion and physical realism — spins, straps and hems that behave like real objects — in clips from 3 to 15 seconds.
Which AI model should I use for a long single-shot fashion film?
Seedance 2.5. It's ByteDance's longest single-shot model, running one continuous take up to 30 seconds, with audio generated alongside the picture.
Which AI model should I use for atmospheric, cinematic campaign film?
Veo 3. Google's model is built for lighting, lensing and mood, in clips of 4, 6 or 8 seconds with native audio — the pick when the shot's atmosphere matters more than its length.
Which AI model should I use for several characters or a character and a product together?
Vidu. Give it up to seven reference images and it composes them into one consistent shot, whether that's a model and a product side by side or several people in the same frame.
Which AI model should I use for legible text or typography on an image?
Seedream 5.0 Pro. It renders native text in 14 languages and holds dense layouts, which most image models aren't built for.
Do I need a separate account with every AI lab Misu uses?
No. Misu holds the licence with each provider and meters usage in tokens. The Studio routes each shot to the model that fits it and keeps your product and character locked across all of them, so you never sign up with ByteDance, Google, Kuaishou or OpenAI directly.
Get started
One brief.
The right model per shot.
Character Lock · Product Lock · Full commercial rights
Brands don't need an account with every lab. Misu keeps the face and the product consistent while the model underneath changes.
BETA · REVIEWED BY HAND · FOUNDING PRICING LOCKED