Inpainting (Gemini)
Repaint just one area of an image. Inpainting runs on Gemini image models (Nano Banana 2.1, 2, 2 Lite and Pro) and appears in the prompt panel when one is selected.
The AI reads the scene before it paints
An inpaint only looks right when the new pixels match what is around them: the light, the material, the style, the lines that run across the edge. Before your request goes out, the Scene Analyzer looks at the area you painted and writes down, in words, what the image model needs to know.
- Your prompt steers the reading. What you type is sent to the analyzer as the instruction, so it reads the painted area for that job. Ask for a red wool coat and it looks at the sleeve seams, the fabric folds and how the light falls on them.
- No prompt? It still reads the scene. Leave the prompt empty and the analyzer works from the image alone: what is under the mask, the mood, where the light comes from and how hard and warm it is, the materials at the edge, the art style and the camera angle.
- What goes into the request: the body part under the mask, the light in words (direction, quality, colour temperature, strength), the materials, art style and perspective, the lines and textures that must continue through the edit, and what the model must not do there. No numbers or colour codes — image models follow descriptive words better.
- What it sees: a crop of the working area up to 1024 px with your mask tinted, a small view of the whole picture so it knows where the area sits, and up to four of your reference images. The full-size original is never sent as is.
- Suggestions, not changes: it also suggests a mode, a preset and which reference to weight. They are marked with a dot and never applied for you.
Which AI, and what it costs
The analysis runs on the assist AI model you pick in Settings (Gemini or ChatGPT) and is billed to your own key.
| Assist AI | Model | About per inpaint |
|---|---|---|
| Gemini | Gemini 3.8 Flash | ≈ $0.01 |
| ChatGPT | GPT-6.1 Sol | ≈ $0.02 |
One call sends two small images (up to six with references) and a short instruction, and gets back a few hundred words. One Nano Banana 2.1 image at 1K, the default model, is about $0.034. Figures are calculated from official prices checked on 4 November 2026: Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens until 31 December 2026, then $1.50 and $7.50 (Google pricing); GPT-6.1 Sol costs $2 and $10 (OpenAI pricing). Thinking tokens vary from call to call.
Presets: start from the right job
Each preset sets the work type, a brush size and a written instruction that fits the job, so the model knows what to protect.
- Generate: Clothing Edit keeps the body, pose and skin tone and fits fabric and folds to the light. Background Edit keeps the subject and matches light direction, colour temperature and perspective. Object Edit fits a new object to the light, perspective and palette around it. Fine Edit makes small fixes that keep skin tone, line quality and style.
- Remove: Object Removal rebuilds texture, pattern, shadow and reflection as if the object was never there. Background Removal keeps the subject's edge clean. Fine Removal clears text, watermarks, blemishes and noise.
- Your own: Save Current Settings stores the mode, work type and a prompt as a preset you can edit or delete.
Options that matter
- Generate / Remove: change or add something inside the mask, or erase it and rebuild what is behind.
- Mask tools: brush, eraser, lasso, polygon and rectangle. Hold Alt to subtract.
- Edit range preview: shows what will actually change (the mask plus a thin blending ring) and the window the model sees.
- Anatomy Awareness: names the body part under the mask, so a shoulder is not painted like an arm.
- Scene (perspective/light): adds the boundary's light, materials, style and continuing lines to the prompt.
- Variation: on Auto, references are reproduced faithfully. Move the slider to let the AI reinterpret more freely.
- Allow 4K requests (cost ↑): only matters when the painted area is larger than the chosen resolution. It requests 4K instead of scaling a smaller result up.