How to Use Generative Fill: Add or Replace Anything in a Photo with a Prompt

What generative fill is

Generative fill is AI inpainting guided by text. You choose a region of a photo and type what should appear there; an image-generation model paints new pixels inside the region while everything outside it stays exactly as it was. It's the same idea Adobe popularised in Photoshop, and it's useful for adding objects (a boat on a lake, a balloon in the sky), replacing things (a plain wall with bookshelves) and extending scenery.

It differs from content-aware fill, which only continues the surrounding texture to hide something. Content-aware fill is faster and better for simple removals; generative fill is for when you want something new.

How it works in the NextG Design editor

The editor runs SD-Turbo, a distilled Stable Diffusion model by Stability AI, directly on your graphics card using WebGPU. Nothing is uploaded. The process has four steps:

1. Context. A region around your selection (with a generous margin) is taken from the photo and resized to the model's working size of about 512 pixels.
2. Pre-fill. The selected area is first filled with content-aware inpainting, so the generator starts from the scene's real colours and layout instead of a blank hole.
3. Generation. The model denoises the selection over a few steps, guided by your prompt, while the area outside the selection is re-imposed from the photo at every step — which is what keeps the perspective and lighting consistent.
4. Blending. The colour difference along the edge of the selection is measured and spread smoothly inward, so the new content meets the photo without a visible seam, and the result is placed on a new layer.

The model is about 2.6 GB. It downloads once, is stored by your browser, and works offline afterwards.

Making a good selection

• Select generously. Include the object's shadow and reflection, and leave some margin. The AI needs room to blend.
• Prefer the lasso to a rectangle. Hard straight edges are the easiest to spot afterwards.
• Size matters. Each result is generated at roughly 512 pixels across the working area, so smaller selections stay sharper. For a large area, fill it in two or three overlapping parts, or tick the AI upscaler option where offered.
• Use Quick Mask or Select › Modify to soften or expand a selection before generating.

Prompts that work

Describe the thing, not the edit. Good prompts are short, concrete and visual:

Adding a medium (“photo”, “watercolor”), the light (“soft morning light”) or the angle (“seen from above”) helps the result match. Very long prompts don't help; the model reads only the first 75 or so words.

Making the result look real

• Generate several variations. Quality varies with the random seed — the third or fourth is often the best.
• Use the layer. The result is its own layer: lower its opacity slightly, paint on a layer mask to soften an edge, or add a Color Lookup or Curves adjustment clipped to it to match the grade.
• Match the grain. Phone and film photos have noise; add a touch with Filter › Add Noise on the generated layer if it looks too clean.
• Check scale and shadows. If an object looks pasted on, the usual culprits are size and missing contact shadows — paint a soft shadow underneath with a low-opacity black brush.

Limits and responsible use

Generative fill invents plausible pixels; it can't know what was really there, and it struggles with legible text, hands and exact likenesses. Requirements: a recent Chrome or Edge with WebGPU and a computer with a graphics card (integrated graphics work but are slower — expect up to a minute per variation). Don't use generated edits to mislead people about real events, and label AI-edited images where it matters, such as news, listings or evidence.

Frequently asked questions