AI Images and Video Inside VULK Projects: Models and What Gets Checked

Verified on October 10, 2026, against the code VULK runs today and its production records.

A generated site needs pictures, and a placeholder box is not a picture. VULK can make the pictures and the motion for your project as part of the build: a hero image in your brand's colours, product shots, a looping video behind a headline. This post explains which models do that work, what sizes and lengths you get, which plans include what, and how a video is checked before it reaches your page.

Images: the model and the formats

The default image model is Gemini 3.1 Flash Image. We chose it after measuring it in August 2026: it tied Gemini 3 Pro at half the price, and it carries the brand's look through a scene without being told to every time.

Every image is described to the model as a production requirement rather than a loose prompt:

• Aspect ratio: 14 options, from square and portrait through 16:9 and 21:9 to extreme banners at 1:8 and 8:1.
• Resolution: 1K, 2K or 4K. A square image is 1024 by 1024 pixels at 1K and four times that per side at 4K.
• Framing: a complete subject, such as a product on its own, or a full scene.

4K images are bought through OpenRouter's Images API, the interface that actually offers that resolution, so a 4K request returns a 4K file.

Images in production

From production records, read on October 10, 2026: between August 24 and October 9, 2026, Gemini 3.1 Flash Image produced 1,004 images across 238 projects, with a median file size of about 1.6 MB.

Video: two places, two defaults

Video appears in two places in VULK, and each has its own default model.

Inside a project, video is made from your project's own images. VULK generates the still first and then animates it with Kling 2.5 Turbo Pro at 1080p, so the subject in the clip is the same one that appears in your pictures. This is the clip you see behind a hero or in a product section.

In Director, VULK's standalone space for images and video, you can generate a clip from text alone. The default there is Seedance 2.5 at 720p, chosen after a delivered five-second clip was verified end to end. Director has its own work view with a rail, a stage and a history of what you made, offers image shot types and accepts a reference image. It sits inside the chat as one of four tabs that share one message box, and 3D objects made there open in an interactive GLB viewer.

You can also pick another engine. Every clip is 5 or 10 seconds long:

In production, between August 26 and October 4, 2026, Kling 2.5 Turbo Pro made 34 project clips across 21 projects, with a median file of about 9.9 MB.

How a video is checked before you see it

AI video fails in recognisable ways: the camera passes through a wall, a person appears out of nowhere, a face melts. VULK checks each shot for exactly those failures before it is used.

A vision model, Gemini 3.7 Flash, receives a contact sheet of frames sampled evenly across the shot and answers a fixed set of questions:

• Do all frames show the same physical space as the first frame?
• Is there an impossible transition, such as passing through a wall, a teleport, a cut or a camera jump?
• Do large objects, people, vehicles or buildings appear that were not in the first frame?
• Is there distortion, such as fisheye warping, melting geometry or extra limbs?
• Is there legible text, signage, a logo or subtitles in the shot?

It answers in a fixed format with the frame numbers it is looking at, and code, not the model, decides what happens next.

Where the files live

Generated images and videos are stored in Cloudflare R2 and served from VULK's own CDN at assets.vulk.dev. Private content is not served by the CDN. Every generated file is recorded against the project and the run that made it.

What you can bring yourself

Generation is not the only source. You can attach your own material to the chat, and VULK treats it as an asset of your project:

• Images in common formats plus GIF, SVG, HEIC, AVIF, BMP and TIFF, stored whole or converted in your browser.
• Video files and GLB models, stored as they are so they can be placed on a page.
• Sounds that the page plays where your request says.
• Font files that the page can declare as web fonts.
• PDF and Office documents, read as text in your browser.

Anything in your project's assets can be pulled back into a request by typing @ in the message box.

Plans

• Images are included on every plan, Builder included.
• Video is not included on Builder, neither inside projects nor in Director.
• Pro, Max, Team, Business and Enterprise include video with no monthly cap on seconds; clips are billed in credits.

When you ask for a video or an image in a message, that request is the purchase: the chat does not stop to ask "are you sure?" before it generates. Plan prices are on the pricing page: Builder $19.99 a month, Pro $39.99 and Max $199, read on October 10, 2026.