LetzAI Tagging System: Instant Models vs. Trained Models

Understanding how Instant Models and Trained Models work with LetzAI's tagging system.

Over time, LetzAI's model training and tagging system has evolved.

This short guide explains how we got here, what changed with the arrival of models like Nano Banana Pro, and what that means for your results today.

1. Trained Models (Old)

Originally, getting consistent results for a person, style, or object required training a custom model.

With early LetzAI versions, this process took a while and was very resource-intensive. That is because during the training process of a model, the visual information you've uploaded gets condensed to build a detailed understanding of the subject.

Once training was finished, the result was a trained personal file that you can reference directly in your prompt by tagging it, for example:

Prompt: @MyModel in a cinematic portrait, soft lighting

The training process could take up to one hour to finish.

👉 For a detailed explanation on how to train V3 models correctly, read our full guide here: V3 Training Guide

2. Instant Models (New)

Nano Banana Pro (and similar modern base models) introduced a different and easier way to achieve consistency.

Instead of relying on a trained model, new base models work simply with one or multiple reference images at generation time. It matches the subject in your prompt directly to those reference images. So it is closer to copying a visual than learning it.

This means that a single high-quality reference image can already provide strong likeness and detail. The biggest upside of this approach is that no dedicated custom model needs to be trained for instant base models.

As a result, the Instant version of your personal models is available immediately on LetzAI, without you having to wait for a training process to finish.

3. How Instant Models work with our tagging system

To keep things simple and familiar, LetzAI integrates Instant Models directly into the existing tagging system.

When you create a personal model on LetzAI, two things happen in parallel:

V3 training starts in the background
A reference dataset is created from your uploaded images that is compatible with the Instant Models

When you combine an Instant Model, such as Nano Banana, with a model tag (for example: @MyModel), LetzAI will automatically merge the prompt with one or multiple images from the reference dataset.

Prompt: Take the likeness of @MyModel and create a cinematic portrait, soft lighting, wearing a different outfit

From your perspective, tagging works exactly the same, even though the technology behind it is different.

Why this is useful

Using an Instant Model with tags means:

✅ You don't need to upload or search for a reference image each time
✅ You can use other public personal models as references
✅ You can keep using the same @tag workflow you already know
✅ You get instant results, even before V3 training finishes

4. Important limitations to understand

Because Instant Models rely on random reference images per tagged model, some trade-offs exist whenever you make use of the tagging system in combination with Instant Models.

Less control over the reference image

Some images that are great for trained models may be poor reference images. For example:

Unusual angles
Extreme lighting
Partial faces
Props, hands, or occlusions

When a weaker reference image is selected, results can look off, exaggerated, or in some cases, miss the likeness entirely.

Multiple tagged models increase this effect

Each tagged model introduces another randomly selected reference image. A single weak reference can already influence the outcome. And several weak references can compound, further reducing output quality.

This does not necessarily mean your model is broken. In most cases, it simply means that the selected reference image was not ideal for that specific generation.

Final note

If you are unhappy with the results when using an Instant Model together with a model tag, you can always revert to "Edit" mode, where you can upload any image without having to tag or train a model.

Try manually uploading the best possible reference image and compare the outcome. You can do so by heading to the EDIT tab in the main menu and selecting the correct Instant Model. In many cases, this alone will significantly improve results.

Instant Models, such as Nano Banana Pro, are powerful, but they behave differently from trained models. Understanding how reference images are selected will help you get better results and avoid unwanted surprises.