Generating images with AI is fast. Generating images that actually look like your brand is the hard part. Every new prompt starts from zero. You paste the same style notes, re-describe the same product angle, reattach the same logo, and the output still drifts. Over a week that is a real amount of time wasted on the same setup.

We added a training image library to RequestDesk so your team can upload reference images once and reuse them on every generation. Pair a training image with a text prompt and nano banana (Google’s Gemini image model) uses it as visual context. The result tracks your brand more tightly than prompt text alone can achieve.

What You Can Upload

PNG, JPEG, or WebP files up to 10 MB and 4096 by 4096 pixels. Typical uploads fall into three buckets:

• Brand assets (logos, wordmarks, signature graphics)
• Product photos (a hero shot, a specific SKU, packaging)
• Style references (an illustration, a texture, a color palette reference)

Each image gets a name, an optional description, and tags so you can filter the library as it grows.

Team-Scoped, Not Per-User

The library is scoped the same way agents and prompts are scoped in RequestDesk. Anyone at your company (and anyone granted access through partnerships) sees the same set of training images. This matches how teams already collaborate in the app. If Alice uploads a brand logo, Bob can reuse it without re-uploading.

No per-user private libraries in version one. If a team wants private images, a separate agent with partnership access handles that case today.

Using a Reference in the Image Generator

Open Image Generation, pick an agent, write your prompt, and click “Choose from training images.” Pick one image from the library and generate. The reference is passed to Gemini as an inline image part alongside the prompt, which keeps the output anchored to what you uploaded.

One honest limitation. The provider dropdown auto-locks to nano banana when a reference image is selected. Gemini is the only provider we have wired for reference-image support today. Flux Pro and OpenAI have reference-image APIs we can integrate later, but for now picking a reference pins the generation to Gemini. The UI tells you this so the dropdown is never a lie.

Saved Prompts Can Bundle a Default Reference

If you save an image prompt that works well with a specific reference image, the library remembers the pairing. Next time you pick that prompt, the reference picker pre-fills with the default. You still get to swap or clear it before generating. Good prompts and good references tend to come in pairs, so this removes one more step from the repeat workflow.

What This Is Not

This is not a diffusion model you train on your images. Nothing is fine-tuned, nothing is stored on Google’s servers beyond a single API call, and there is no opaque “model of your brand” being built in the background. Each generation is stateless. The reference image is sent with the prompt, Gemini generates, and that is it.

Usage is tracked per training image so you can see which references are pulling their weight and which ones were one-off experiments.

What Is Next

Two expansions are obvious from here. First, extending reference-image support to Flux Pro and OpenAI’s gpt-image-1 so the provider dropdown works with references instead of locking. Second, multi-reference generation (more than one image per prompt) for cases where you want to combine a brand logo with a product shot. Both are on the backlog.

For now, the feature is live at app.requestdesk.ai under “Training Images” in the sidebar, with the reference picker inside the existing Image Generation tool.