The video model is becoming a studio

Duration. Up to 30 seconds per generation; extend twice

Sound. Joint audio-video generation

Working mode. Multimodal reference, timestamp edits, white-model & green-screen

The product page states that users can create videos up to 30 seconds in a single generation, with the option to extend twice for richer storytelling, alongside smoother motion and more realistic visuals. The launch article adds that single-pass generation grows from 15 to 30 seconds and supports multi-round extensions, so a story can unfold through setup, development, turning points, and resolution rather than stretching one moment.

Reference and editing form the second axis. Official material says the model understands reference video more precisely—intention, framing, and cinematic language—beyond simple motion transfer. The launch article lists up to 30 images, 10 video clips, and 10 audio clips in one pass, plus clay render, motion, and creative references. Editing includes timestamp-level audio-video control, green screen, camera perspective, and reference-based edits for film and advertising workflows.

For professional production, the product page highlights white-model control, green-screen editing, camera movement, and performance blocking. The launch article says Seedance 2.5 is rolling out on Jimeng AI, Doubao Pro, and other creative products. This brief records those public claims only; it is not an independent benchmark.

Key points: Up to 30 seconds per generation, with the option to extend twice (product page); Joint audio-video generation; multi-round extension for longer narratives (launch article); Up to 30 images, 10 videos, and 10 audio clips as references (launch article); Timestamp editing, white-model control, green-screen and camera edits (official pages).

Links: Product page, Launch article.

Duration. Supports extension; limits depend on product surface

Sound. Native audio (dialogue, ambience, music, effects)

Working mode. Video generation in Gemini and Google Flow

The official capability story focuses on greater realism and fidelity grounded in real-world physics and audio, stronger prompt adherence, and improved creative control that now spans audio. DeepMind’s own sample prompts describe shot scale, performance, city murmurs, a hip-hop bed, and spoken dialogue in one instruction—audio is authored with the picture, not attached later.

The DeepMind page points creators to Gemini and Google Flow. Gemini offers a direct conversational entry, while Flow is organized around shots and longer creative sequences, so the practical choice depends on the complexity of the project.

Relative to silent clip generators, Veo 3.1’s public difference is audiovisual unity: dialogue, ambience, and rhythm can arrive with the image. Exact duration caps, resolution options, and regional availability follow each surface’s current documentation; this article does not invent a single public limit where the pages do not state one.

Key points: DeepMind positioning: leading video model for filmmakers and storytellers; Emphasis on native audio, physical realism, prompt adherence, and creative control; Creative products include Gemini and Google Flow; Official demos write ambience, music, and dialogue into the same prompt.

Links: Google DeepMind.

Duration. Gemini overview states about 10-second videos

Sound. Native audio generation

Working mode. Multi-turn editing with text, image, and video

The Gemini overview lists create ~10-second videos, native audio generation, turn photos into video (up to 5), video-to-video editing, and multi-turn editing. In the Gemini app, Omni is described as replacing the previous Veo 3.1 path. DeepMind’s model page adds the framing “create anything from any input—starting with video,” with natural language edits that each build on the previous state.

The Google Blog introduces Gemini Omni Flash as the first model in the family, rolling out to the Gemini app, Google Flow, and YouTube Shorts. Users can combine images, audio, video, and text as input and generate video grounded in Gemini’s real-world knowledge. Across turns, official copy stresses character consistency, physics that hold up, and a scene that remembers what came before.

The public differentiator is interaction: swap a background, change wardrobe, adjust lighting, or replace a character through chat instead of restarting from zero. This brief follows those official pages only and does not treat demos as controlled comparative tests.

Key points: Gemini overview: ~10-second videos, native audio, up to 5 photo references; Video-to-video editing and multi-turn conversational edits; Text, image, audio, and video can be combined as input; Rolling out to Gemini app, Google Flow, and YouTube Shorts (Google Blog).

Links: Gemini overview, Google DeepMind, Google blog.

Duration. Short-form generation; official page showcases 2K + stereo

Sound. Native stereo audio

Working mode. Unified multimodal-context understanding and generation

The official example is concrete: borrow Hitchcock-style camera movement from a video, have a character from an image sing, and match the performance to a separate audio track. The creator describes how the materials relate instead of handling each reference in a different tool.

Earlier products often separated image animation, first-and-last-frame control, subject, motion and voice references, and editing into different workflows. H3 aims to keep them in one creative context, where multi-shot video, sound, references, and natural-language revisions can continue together.

The page also highlights 2K output and native stereo sound, with use cases spanning film titles, product showcases, animated posters, and advertising. For creators, the useful question is whether different kinds of source material can be understood and used together.

Key points: Unified text, image, video, and audio context (official research page); Showcases 2K output and native stereo sound; Relationships among references described in natural language; Suited to film titles, product showcases, animated posters, and other mixed-media work.

Links: MiniMax research.

Duration. Up to 15 seconds

Sound. Native audio with character-level lip sync

Working mode. Multi-shot storytelling, element consistency, multilingual lip sync

The product page lists what creators can make: multi-shot cinematic shorts with dynamic camera movement, localized ads with characters speaking different languages and accents, and up to 15 seconds of high-fidelity continuous action. What’s new emphasizes cinematic audiovisual synergy, element and character consistency, and native text rendering inside 15-second sequences for ads, product demos, and commercial content.

The VIDEO 3.0 Omni guide contrasts O1 with the new model: text- and image-to-video gain native audio and multi-shot support; video element reference and element voice control are added; duration rises from up to 10s to up to 15s. Images, videos, elements, and text are all treated as prompts, and the model can lock multiple characters or items across complex group scenes.

Official marketing also presents cinema-grade native 4K as part of the series. For deliverable short-form work, the stated priorities are stable subjects, accurate lip movement, readable text, and continuous shots—not a single spectacular frame. Capability claims here come only from the Kling product page and Omni user guide.

Key points: Up to 15 seconds per generation; Omni guide notes the rise from 10s to 15s; Native audio with character-level lip sync; multi-person and multilingual marketing; Multi-shot storytelling, element/character consistency, video elements and voice control (Omni); Product page emphasizes native text rendering and cinema-grade native 4K.

Links: User guide, Product page.