> ## Documentation Index
> Fetch the complete documentation index at: https://docs.wji.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Reference images

> Add references, label them by type, and understand the per-model limits.

References are how you keep a character looking like the same person across a hundred shots. They
matter more than prompt wording for consistency.

## Six ways to add one

<Steps>
  <Step title="Pick a file">
    Click to browse. Multiple selection works.
  </Step>

  <Step title="Drag or paste">
    Drop files onto the prompt box or paste them. Only images are accepted this way.
  </Step>

  <Step title="From a dataset">
    Choose from your workspace ideas, grouped by character, location, outfit, prop, and style.
  </Step>

  <Step title="From IP assets">
    A separate tab for virtual character assets.
  </Step>

  <Step title="From a result">
    Hover any history result and add it as a reference directly.
  </Step>

  <Step title="Recently used">
    The last six references you used stay one click away.
  </Step>
</Steps>

## Reference types

<Frame caption="Reference images with their types set">
  <img src="https://mintcdn.com/wji/1nwL1jaWRdmAhYLq/images/image-generation/reference-block-01.png?fit=max&auto=format&n=1nwL1jaWRdmAhYLq&q=85&s=825897857c3d3ec27e2924f6f2d098d1" alt="Reference block with several cards, each labelled by type such as character, location, and style" width="1600" height="900" data-path="images/image-generation/reference-block-01.png" />
</Frame>

Every reference carries a type. Set it by clicking the card.

| Type          | Use for            |
| ------------- | ------------------ |
| **Character** | People             |
| **Location**  | Places and sets    |
| **Outfit**    | Clothing           |
| **Prop**      | Objects            |
| **Style**     | Look and treatment |
| **Layout**    | Composition        |
| **Other**     | Anything else      |

References pulled from a dataset inherit their type automatically, which is a good reason to keep
your idea types accurate.

Each reference can also be named. Names default to `image1`, `image2` and so on, and are what you
address with `@` mentions in the prompt.

## Limits

**The maximum number of references is set by the model, not by WJI.** When you exceed it, new
references are added at the end and the oldest ones are dropped from the front.

Vertex-protocol models cap a single reference image at 20 MB.

## Reference video and audio

In video mode you can also add reference video and audio.

| Constraint      | Value                                       |
| --------------- | ------------------------------------------- |
| Format          | MP4 or MOV                                  |
| Default maximum | 3 clips                                     |
| Duration        | 2–15 seconds, extended to 30 on some models |
| File size       | 50 MB up to 15s, 200 MB for 15–30s          |
| Dimensions      | 300–6000 px per side                        |
| Aspect ratio    | Between 0.4 and 2.5                         |
| Pixels          | Up to 1920×1080                             |

<Warning>
  If WJI cannot read a video's duration, the size limit drops to 50 MB regardless of length.
</Warning>

## Weights

There is a single **reference weight** slider applying to all references at once, not one per image.
Midjourney-family models add separate image weight and style reference weight controls in the
model-specific settings.

## On the canvas

The Infinite Canvas does not use an upload list. References come from **connections** — you wire an
image, video, or audio node into a generation node and it becomes a reference. See
[Connections and references](/infinite-canvas/connections-and-references).