AI Images & Video

Setting Up AI Images and Video

The platform can generate pictures, and short video clips with sound, from a written description, using Google's Gemini. Each result is saved as a file owned by the person who asked, and every call to Google is a priced run on AI Usage. Because it spends money on whatever a person types, it is switched off until you decide who may use it and how much.

Where to find it

Architect Panel → Configuration:

• AI Settings — the Gemini (Google) section for the key, and the AI images and video section

Architect Panel → Integration & Connections:

• AI Tool Access (MCP) — lists the image and video tools once they are served

Architect Panel → Automation:

• Tasks — Background Jobs, which saves finished videos

Where people use it

• From an AI assistant over MCP: the tools generate_image, start_video and get_video. See MCP Server.
• From a custom feature: a developer can add the platform's image and video dialog to a screen of your app, or call generation from the app's own code.

There is no standard platform screen with an image generator on it yet, so on an installation with no MCP clients and no custom feature, switching this on makes nothing visible.

Switching it on

1. On AI Settings, enter a Gemini API key in Gemini (Google) and Save.
2. In AI images and video, set Who may generate: Nobody (off), Architects, Administrators, or Everyone signed in.
3. Optionally list security group IDs, one per line, in Also these security groups. Their members may generate at any level except off.
4. Set Images per person per day and Videos per person per day. 0 means none.
5. For MCP use, switch on Offer to MCP clients.
6. Press Save on the section.
7. In Tasks, switch on Background Jobs if anyone will make videos.
8. Generate one image, then find it on AI Usage with its cost.

Who may generate

One setting covers every route: MCP, the dialog and AI tools. Architects are allowed at every level except off. Administrators adds administrators, and Everyone signed in opens it to all users. The daily allowances apply to each person separately: images count each successful image, and videos count every video started that did not fail, since a running video is already being paid for.

The other settings

• Default image model and Default video model: blank uses the built-in defaults. Image role model and Video role model override them for the platform's own routes.
• Images per browser request: how many pictures one request in the dialog may make, 1 to 4.
• Image models the browser may pick and Video models the browser may pick: one per line. Empty offers every built-in model not past its shutdown date.
• Send Google a per-person safety identifier: on by default. Google receives a one-way code per person, so misuse is attributed to one user rather than the whole installation.
• Delete Google's copy of a video: on by default, once the video is saved here.
• People in videos (Veo): for the older Veo models only. In the UK, EU and Switzerland only allow_adult is accepted.
• Abandon an unfinished video after (s): default 900.
• Largest reference image, Largest request and Largest video download: size limits in bytes.
• Price overrides and Unknown model capabilities: JSON, for a negotiated price or a model the platform does not yet know.

Cost and credit

Each image and each video start is a run on AI Usage with its estimated cost. Where AI billing is on, the tenant's credit is checked before each image and each video start, and the call is refused before anything is sent if the credit is used up. Checking on a video and downloading it are never refused, so a video that has been paid for is always collected.

What goes wrong, and how to tell

• "AI image and video generation is switched off on this system": Who may generate is off.
• "AI image and video generation is not set up on this system": there is no Gemini key.
• "Your account is not allowed to use AI image and video generation": the person is outside the level and the listed groups.
• "You have used 25 of your 25 AI images for today": the daily allowance is spent; it resets at midnight.
• Nothing about images appears to users: expected without an MCP client or a custom feature.

Worked example

A communications team uses an AI assistant connected over MCP. The architect adds a Gemini key, sets Who may generate to Architects, adds the communications team's security group under Also these security groups, sets five images and one video per person per day, and switches on Offer to MCP clients. A team member asks the assistant for a banner image; it appears in the assistant and is saved as a file, and the run on AI Usage shows its cost.

Recommendations

• Start with a narrow audience: Architects plus one named group.
• Keep video allowances low: a video costs many times an image.
• Switch on Background Jobs before allowing video.
• Leave the safety identifier and the deletion of Google's copy on.
• Review AI Usage weekly for the image and video features once it is open.

Image and Video Models

Which Gemini model makes an image or a video decides the sizes, shapes and options on offer, and some models have published shutdown dates. Videos also behave differently from images: they take minutes, run in the background, and need something to collect them. This article covers both, and what each failure message means.

Where to find it

Architect Panel → Configuration:

• AI Settings — AI images and video: default models, the models offered, and the video limits

Architect Panel → Activity:

• AI Usage — each image and video run, with its warnings and cost

Architect Panel → Data:

• Background Jobs — each video in progress and how it ended

Image models

• Nano Banana 2 (Gemini 3.1 Flash Image), gemini-3.1-flash-image: the default. The widest range of sizes and shapes.
• Nano Banana Pro (Gemini 3 Pro Image), gemini-3-pro-image: sizes from 1K to 4K.
• Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image), gemini-3.1-flash-lite-image: 1K only.

Each can be given reference images to work from, within the size limits set on AI Settings. Google's older image models have been shut down and are no longer offered.

Video models

• Gemini Omni 1.1 Flash, gemini-omni-1.1-flash: the default. Landscape or portrait, from 360p to 4K, with sound. It chooses the clip's length itself, a few seconds up to about ten.
• Veo 3.1, Veo 3.1 Fast and Veo 3.1 Lite: preview models that Google shuts down on 22 October 2026. Every call to them carries a warning naming the date and Google's replacement, and from that date they drop out of the models offered.

Choosing which models are offered

1. On AI Settings, set Default image model and Default video model, or leave them blank for the built-in defaults.
2. To restrict what a person may pick in the image and video dialog, list model ids one per line in Image models the browser may pick and Video models the browser may pick.
3. For a model the platform does not know yet, describe its sizes and shapes in Unknown model capabilities. Without that, its options are sent unchecked and each run carries a warning saying so.
4. Save, and generate one image or video with the new model to check it.

How a video is made

Starting a video returns at once with a job; Google then takes anything from a few seconds to several minutes. Something has to keep asking Google and save the clip when it is done:

• The Background Jobs task, when it is switched on and the task engine is running.
• The person's own screen or MCP client, by checking the job. A video started in the dialog or over MCP is finished by whichever comes first.

If the worker is not running, the dialog warns before the video is started and asks the person to keep the window open, and an MCP client is told to keep calling get_video. A video requested from an app's own code or an AI tool is refused instead, because nothing would collect a video that Google would still charge for. A video unfinished after Abandon an unfinished video after (s) is given up.

Once a video is saved, Google's copy is deleted while Delete Google's copy of a video is on.

What the messages mean

• "Google declined to make this image" (or video): Google's safety checks refused the description. Reword it.
• "The reference images are too large.": over Largest reference image or Largest request.
• "The video took too long and was abandoned.": it passed the abandon limit.
• "The AI provider is busy or a spend limit was reached.": Google is busy, or the Google account is over its limit. Try later.
• "Your organisation's AI allowance has been used up": AI credit refused it before anything was sent.
• "Videos cannot be made in the background right now": switch on the Background Jobs task.
• "Image generation is not set up correctly on this system" (or video): the key was refused, the model id is unknown, or a media role names a provider other than Gemini. Check the key and model ids on AI Settings.

A cost to watch for

If a video is started while the worker is off and the person closes the window before it finishes, Google still makes and charges for it, but the platform never saves it and may record no cost for it on AI Usage. Switching on Background Jobs removes the problem.

Worked example

An installation set up in the summer has Default video model set to a Veo id. AI Usage starts showing warnings on every video run that the model shuts down on 22 October. The architect clears Default video model, so the built-in Gemini Omni default applies, generates one test clip, and confirms the warning has gone before the shutdown date arrives.

Recommendations

• Move off Veo before 22 October 2026, or leave the default video model blank.
• Read run warnings on AI Usage: shutdown notices appear there first.
• Keep Background Jobs on wherever video is allowed.
• Restrict the models offered if cost per image varies more than you want people choosing.