Caira Adds Generative Video Editing – Gemini Omni 1.1 Flash, Runway Aleph 2.0, and 720p or 1080p Output

Camera NewsAI

by Alexandra ThompsonSeptember 7^th, 2026

Subscribe to News - Cameras

Camera Intelligence is developing its second-generation camera platform, Caira. The interchangeable-lens Micro Four Thirds camera attaches to an iPhone via MagSafe, using the phone as its screen and interface. Camera Intelligence says it records up to 4K/30fps and 1080p/60fps, using a 4/3-type Sony Quad Bayer sensor, a Qualcomm Snapdragon processor, and a Google Edge TPU.

Caira follows the earlier Alice Camera project, which also paired a Micro Four Thirds camera with a smartphone-based interface. We previously spoke with CEO and co-founder Vishal Kumar about Alice and the thinking behind that project, and in case you missed it, click here for the full interview.

What the camera records, and what it does not

Nothing generative happens at the point of capture. Camera Intelligence states that Caira’s on-camera image pipeline remains non-generative, so the file written when you press record is a conventional recording from the sensor. The company says original recordings are not modified and generative results are saved as separate files.

Editing is initiated in the Caira app, while the press release attributes the generative processing to Google’s Gemini Omni 1.1 Flash and Runway’s Aleph 2.0. Camera Intelligence also states that customer footage is never used for model training, “neither by us or by Google’s API.” That reference to an API suggests the generative processing happens off-camera rather than on Caira’s Edge TPU, although the company does not explicitly say where the compute runs. Generation times are stated as one to three minutes depending on resolution.

The four launch templates

Generative Video Editing launches with four templates: Relight, Reframe, Camera Motion, and Set Extension. Editing takes place through a curated template library developed by Camera Intelligence with feedback from its community. The company says the templates can perform color grading, simulate drone-style sweeps, and create other camera paths. Users can then refine a template with typed or spoken natural-language instructions, which the company says are interpreted within Caira’s ethics guardrails.

It is worth being precise about what these edits actually do to the recorded image. Set Extension and Camera Motion both introduce visual information that was not captured by the sensor: an extended background must be generated, and a camera move that was never filmed requires the model to synthesize new views of the scene. Relight similarly goes beyond a conventional exposure or color adjustment by generating a different lighting treatment for the recorded scene. In one press image, a Day to Night template is described as turning daytime footage into night “with any practical lights in the scene coming on,” while the prompt field reads “Turn the scene into golden hour.” If those practical lights were not illuminated in the original footage, their appearance in the result is generated rather than recorded.

Reframe and color grading are less clear-cut. Reframe could simply mean recomposing within the existing image area, which would amount to a crop, or it could involve generating image information beyond the original frame. The material provided does not specify which approach Caira uses. The same ambiguity applies to color grading: the press release says the templates can perform color grading, but it does not explain whether this is a conventional image adjustment or a generative reinterpretation of the frames.

What “physics-aware” is being claimed

Camera Intelligence says these models differ from earlier generative approaches because they do not treat each frame independently. According to the press release, the models build a persistent internal understanding of a scene’s geometry, lighting, materials, and motion. The company describes Gemini Omni as “physics-aware” and says it can remember a scene across an entire clip so that relighting or set extension remains consistent as people, objects, and camera movement change from frame to frame.

The company provides demonstration footage showing effects such as Golden Hour and other generative treatments applied across moving shots. These examples support the claim that the system can maintain a treatment over time, although they do not amount to a controlled technical test of the underlying “physics-aware” model behavior.

Output resolution is the number to watch

Generative edits are delivered at either 720p or 1080p, while Caira itself records up to 4K30. That means footage sent through a generative pass returns at a lower resolution than the camera’s maximum capture format. For filmmakers working toward a 4K final deliverable, that is a practical limitation rather than just a specification detail.

Camera Intelligence positions the feature more as a tool for rapid previsualization and prototyping than as a replacement for full production. One example given by the company is a user shooting a spec sequence in the morning, presenting several treatments the same day, and then shooting the final production later with a full crew.

There is no text-to-video function; every generative edit begins with footage captured on Caira.

Labeling, watermarking, and what SynthID covers

Generated results are watermarked with Google’s SynthID, which Camera Intelligence says is used to identify AI-generated content. Original recordings are not watermarked.

The company does not explain whether the same SynthID treatment applies equally to outputs generated with Runway Aleph 2.0, how well the watermark survives re-encoding or further editing in an NLE, or whether the files also carry any human-readable label, metadata flag, or C2PA Content Credentials. Those details matter if the goal is to make the distinction between recorded and generated footage clear after the file leaves the Caira app.

Camera Intelligence’s Ethical AI Charter

Alongside the announcement, Camera Intelligence published an Ethical AI Charter outlining how it says generative features will be handled. The company states that customer footage and images are never used for model training by Camera Intelligence or through Google’s API. It also says its guardrails prohibit alterations to an individual’s skin tone, ethnicity, or fundamental facial features, and that it uses pre-vetted commercial models while planning to adopt licensed, opt-in models as they become available.

Camera Intelligence also says its AI systems follow Google’s Generative AI Prohibited Use Policy and that it applies what it describes as ethical system prompts and safeguards. These commitments set the company’s intended boundaries, but they do not, by themselves, show how those safeguards perform in practice.

Kumar also acknowledged the broader debate around AI in photography and filmmaking:

We understand there are strong emotions around AI in photography and filmmaking right now, and we want to be cognisant and respectful towards artists, their craft and community.

Vishal Kumar, founder of Camera Intelligence

He added that traditional filmmakers and VFX artists may prefer their existing workflows, while Camera Intelligence is targeting solo creators and marketers who may not have the budget for more conventional production or post-production work.

The company says the feature was shaped in part by a survey of 111 Caira backers, 77% of whom said generative video editing would be useful to them.

For filmmakers, this is still largely unknown territory. Generative tools like this begin to move decisions that traditionally belonged to production: lighting, framing, camera movement, even parts of the environment into post. That does not remove the need to shoot carefully, but it changes what “capturing the shot” can mean. Some filmmakers will see that as useful flexibility; others will see it as a much more fundamental shift in authorship, craft, and the relationship between what was filmed and what finally appears on screen.

Price and availability

Generative Video Editing will be part of Caira Studio, a paid software tier launching this fall. The subscription will also include Generative Image Editing and one additional feature that has not yet been announced. Pricing has not yet been confirmed.

The feature will initially be available only in the US and UK. A private beta opened on September 2 for existing Caira customers and five working filmmakers selected through an application process.

Caira itself remains priced at $999, with Kickstarter units continuing to ship in monthly batches.

So where do you draw your own line between an edit and a generation, and does a separate file plus an invisible watermark do enough to mark it? Let us know in the comments below!