Render a subject with a transparent background Transparent background

Get your subject on a real transparent background, ready to drop onto anything.

5 stepsNothing to prepare

The same tutorial, three subjects

Step by step

Skill level

  1. Prompt

    1 · Describe one isolated subject

    A cut-out is only ever as clean as the line between the subject and what is behind it, so ask for one thing rather than a scene.This example asked for:
    a small brass pocket compass on a braided leather cord, engraved bezel, scratched glass, studio product photography, sharp focus, even lighting, isolated subject, plain background
    The negative keeps a setting from creeping in:
    background, scenery, lowres, blurry, watermark, text
    A landscape gives the model nothing obvious to cut around.

  2. Workflows

    2 · Turn on Transparent background

    This is the whole technique, and it is one switch, , in the Workflows card.It selects Layer Diffuse, which paints your subject onto a real alpha channel instead of onto a backdrop you would then have to cut away by hand.

  3. Model & style

    3 · Pick the model

    Transparent output needs a Stable Diffusion 1.5 or an SDXL model, and the panel says so if you pick anything else.This example used AlbedoBase XL (SDXL), an SDXL one, which is why it asks for 1024×1024.

  4. Size

    4 · Set the size, and leave the rest

    1024×1024 is the size an SDXL model is trained at.Nothing else about this request differs from a plain text-to-image one: the Sampling card stays at k_euler_a, 30 steps, guidance 7.5, seed 1234567890.

  5. 5 · Send it, and keep the alpha

    Press Generate, and when the image arrives, save it as webp or png.Most of this image is fully transparent, and jpeg has no alpha channel at all: saving it that way flattens every transparent pixel onto a colour and throws the cut-out away.Expect anything from a few seconds to a few minutes. This run took about 127 seconds and cost 45 kudos.

Where next

You can now ask for a cut-out instead of a backdrop. The last of the workflows hides a code that a phone can still read inside a picture, and it teaches something about prompts: what the picture is asked to supply when a ControlNet is already carrying the structure.