Media Library and AI Images

The Media Library at Marketing > Media (/dashboard/marketing/media) is where every image and video in your account lives: uploads, AI generations, Design Studio exports, imported photos. It is what the social composer, the Design Studio, the workflow builder and the AI assistants pick from when they need a visual. This page covers the rules that apply to everything in it.

​

Which model draws the picture

As of 2026-09-12 the default image model everywhere in Hiveku is GPT Image 2.5 (OpenAI’s gpt-image-2.5-flare, released 2026-09-08): the Media Library’s generate action, the Design Studio chat, the Generate Image and Generate Image Set workflow nodes, the Visual Editor, the article images below, the agents, the Claude Code plugin and the API. It is better than the previous default at about half the latency, and it keeps a referenced subject intact on an edit, which is what matters when a real product photo is being restaged. Every generation is priced at what it actually cost, so the credit ledger charges the real figure rather than an estimate. Two things to know about it:

• A provider outage does not dead-end a generation. When the OpenAI lane fails, the same request renders on Gemini 3.1 and the result says which model rendered. Gemini 3.1 and GPT 5.4 stay selectable when you want them.
• A premium edit model exists. gpt-image-2.5-sunburst is the pick for an edit that must hold the referenced subject exactly across several rounds; ask for it by name in chat, or pass model: "gpt-image-2.5-sunburst" on the API or through the plugin. The flux, flux-pro and recraft models are a separate lane with their own per-image billing, exact frames, and seed support.

​

Brand-aware image generation

Every AI image generation applies your active brand guide by default. Your own prompt is stored untouched as the image’s title and metadata; what the model additionally sees is the brand image profile described below. This holds wherever images are generated: the Design Studio chat, the Generate Image, Generate Image Set and Generate Video workflow nodes, article images, and any assistant working through the Claude Code plugin or the API. The result is honest about what happened:

• Brand applied: the generation records that the active brand guide reached the model, and brand_reference_ids lists which reference images went with it.
• No active brand guide: the image is generated unbranded, and the result says so. Activating a guide under Marketing > Branding is what turns brand context on; nothing else has to change.
• Brand guide could not be read (a database problem on Hiveku’s side): the generation is refused before anything is spent, instead of quietly producing and billing an unbranded image. Retry in a moment, or opt out for that run.
• A reference image could not be fetched: the generation still runs and the result names the reference that was skipped (brand_references_skipped), so a missing logo is never a silent omission.

To generate without brand context for one run:

• In chat: say so. “Generate this without brand styling” is enough.
• In a workflow: set Apply Brand Context to No on the node.
• Through the API or the Claude Code plugin: send use_brand: false on the request.

​

The brand image profile

Until September 2026 an image model asked to stay on-brand was told the brand name, the industry, three colors, two fonts and 200 characters of voice. The brand guide knows far more about how the brand looks, and as of 2026-09-12 all of it reaches every branded generation as one brand image profile: The profile also carries reference images, attached to the request as pictures rather than words: the primary logo first, then up to three images from the Media Library that you have marked as approved brand imagery (see the tag below). Only the primary logo travels; the dark, secondary, icon and monochrome variants stay home, because handing a model five logos is how the wrong one ends up on the picture. An assistant can read the profile back exactly as a model will receive it (GET /api/olympus/marketing/brand/image-profile, the brand_image_profile_get tool), which is the fastest way to see what a vague guide is producing. An empty photography line in that read is fixed in the brand guide, not by adding adjectives to every prompt.

​

Approved brand imagery: the brand_reference tag

To put your own look on every generated picture, tag Media Library images brand_reference:

​

Images for an article

A blog post or article gets its pictures as one step rather than a featured image alone: a hero, and one image under each major section where a picture adds information the words do not. See Images for the piece for how the Content department, the editor’s Generate images action and the API run it, what it costs against the allowance below, and how the alt text is written.

​

Your monthly image allowance

Each plan includes a number of AI image generations per month: 5 on Free, 10 on Starter, 100 on Standard, 150 on Premium. See Billing for the full plan table. What draws from it:

• Each generated image. An image set counts one per image it produces, and an article images run counts one per hero or section image it makes. A generation that fails is refunded.
• The composed reference still when a video is generated in compose mode.
• Each upscale (see below).

Importing by URL, transforming an existing asset, exporting a design and stock photo searches do not count. Before a large batch, ask your assistant how many generations are left this month. The read reports what has been used, the plan limit, what remains and when the period resets, and it is exact: it reads the same counter the next generation is gated on. If the count cannot be read, the assistant is told so rather than being handed a zero.

​

Importing an image from a URL

There are two ways to bring an outside image into the library, and they are not the same:

• Register the URL: the library stores a pointer to the other site’s file. If that site removes or moves the image, your asset breaks with it.
• Import by URL: Hiveku downloads the bytes, checks they really are an image, stores them in your own media storage and registers a new asset that outlives the source. The original URL is kept on the asset for reference.

Import by URL is what you want for anything you intend to keep using: a product photo from a supplier site, a press image, a photo the client emailed a link to. Ask the assistant to import it, optionally with a title, tags and a folder. The limits: the URL must be http or https, the file must be an image (SVG is refused because it can carry scripts), the download is capped at 25 MB, and the copy counts toward your plan’s media storage.

​

Transform: crop, resize, re-encode

Simple pixel operations on an image already in the library, done on Hiveku’s side so nothing needs to be downloaded and re-uploaded: Operations run in that order when combined. Output is capped at 4096 px per side; the source can be up to 25 MB and 24 megapixels. The result is a new asset titled after the original with an “(edited)” suffix, linked to its parent so you can trace where it came from. Transforms are free: they do not touch the monthly image allowance. Only your media storage is affected.

​

Upscale

AI upscaling for print, large-format and retina uses. The source stays as it is; the upscaled result lands as a new asset with an “(upscaled)” suffix, linked to its parent.

• Scale: 1x to 4x, 2x by default.
• Cap: the output must stay within 32 megapixels (source pixels multiplied by the scale squared). A 2000 x 1500 photo is 3 megapixels, so 3x (27 megapixels) is fine and 4x (48 megapixels) is refused up front. Nothing is submitted or charged for a job that cannot fit.
• Cost: one image generation from your monthly allowance, plus the upscaler’s own charge of $0.03 per output megapixel deducted from your AI credits. A 16 megapixel result costs$0.48. Only the source’s real dimensions, measured at the moment of the request, are used for the cap and the price.

​

Every operation makes a new asset

An asset’s bytes are immutable. Its file, dimensions, duration and generation metadata cannot be edited once it exists; what you can change is the title, alt text, caption, description, tags and folder. Transform, upscale, import and every Design Studio export add a row rather than replace one, which is why two versions of the same image can sit side by side and why nothing you or an assistant does can silently damage an asset a live post or page is already using. Delete what you no longer want.

​

What’s Next?