Gemini Omni, announced at Google I/O on 19 May 2026, is the named example: text, image, audio and video in; editable video out. Omni Flash followed on 30 June 2026 as the conversational editing tier. Seedance 2.0 sits in the same family of idea with unified multimodal input (text, image, video, audio, up to 12 references) even if the marketing word is not 'omni.' The older pipeline was a still, then a silent clip, then a voice, then captions. Any-to-any is the claim that those steps can be one sample.
Multimodal is the weaker label. A model can read images and still be unable to edit them, or generate strong video and treat audio as an afterthought. Any-to-any is the interface claim: you may start from whichever artefact you have and ask for whichever artefact you need. In practice the quality is still uneven by direction. Image-to-video is more mature than video-to-video. Native audio is better at ambience than at scripted dialogue. The catalog row is where you find which directions a given model actually runs.
For a creator the change is the brief, not the magic. You stop writing a prompt that has to invent every input you already own. You attach the product stills, the voice take, the previous shot, and you describe the change. Multi-reference is the stills version of that habit. First-last frame is the two-still version. Any-to-any is the version that also takes a clip and a line of speech. It does not retire specialists. It retires the assumption that every job starts from a blank text box.
In practice
- Start from the artefact you already have — a still, a take, a track — and describe the change, rather than regenerating the world from text.
- Check which input slots the row actually exposes. 'Omni' in a keynote is not a slot on the form.
- Keep a specialist for the job the omni model is mediocre at. Mixed inputs are not equal outputs.
The mistake to avoid
Hearing 'any-to-any' and retiring the image model, the TTS row, and the caption pass. One sample can do those jobs. It does not yet do all of them well.
Where you will run into it
- Gemini Omni Video — Google Gemini Omni multimodal video generation. Accepts a prompt plus optional reference images, source video clips, character IDs, and audio IDs. Supports 720p / 1080p / 4k output at 4-10 seconds (16:9 or 9:16). Quota: images + videos*2 + character_ids <= 7.
- Seedance 2.0 — Seedance 2.0 generates Hollywood-grade cinematic videos from text prompts with native audio-visual synchronization, director-level camera and lighting control, and exceptional motion stability
Related terms
Multimodal model
A multimodal model handles more than one kind of data — text, images, audio, video — inside a single system rather than bolting separate tools together.
Native audio
Native audio meaning: a video model that generates soundtrack (dialogue, effects, ambience) in the same pass as the picture.
Multi-reference
Multi-reference generation conditions a clip on several stills at once — a stack of identities or products, not one picture and a prompt.
Reference-to-video
Reference-to-video means the model builds a clip from separate images of a subject it can place anywhere. Those images are an identity, not the first frame.
Diffusion model
A diffusion model generates by starting from random noise and removing a little of it at a time until a picture or clip is left behind.
The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync, in your browser or on your phone.
Free on the web and on iPhone, one account.