FLUX 3 Image Klein prompting guide

1. Write the whole picture

Subject → action → setting → light → camera → style. Compare:

Short:  a cat on a sofa
Better: A close-up photo of a ginger cat curled up asleep on a mustard-yellow
        velvet sofa, late-afternoon sun from a window on the left, shallow depth
        of field, 50mm lens, warm natural colors.

Too lazy to write long prompts? Use the prompt upsampler (flux-3-image-klein-pup, built into the reference CLI and the API's prompt upsampling).

2. Text in the image

FLUX 3 Image Klein can render text. Put the exact words in quotes and say where they go: a vintage enamel sign that reads "OPEN 24 HOURS" above the door. Check spelling in the result — the model card says rendered text can still come out wrong.

3. Editing one image

State the change and what stays: Make it night with heavy snow; keep the tram, the street and the camera angle the same. Through the API, pixel restoration keeps untouched areas pixel-identical.

4. Several reference images

You can pass up to 10 images. Refer to them by their position so the model knows which is which — for example the woman from image 1 wearing the jacket from image 2, standing in the room from image 3. With bounding boxes, the input is written as <ref_image_0>.

5. Control the layout

When position matters, end the prompt with bounding boxes — see the bounding-box guide.

Don't

• Don't put real people in fake or intimate scenes, or copy someone's likeness without consent — the license forbids it.
• Don't expect factual accuracy: the model draws, it doesn't look things up.