Fix one part of a picture and leave the rest alone: the inpainting method, plus where it lives in each tool

Check it is an inpainting job

Inpainting is for pictures that are right except for one part: a hand, a sign, a stray person, a label, the sky. If the composition or the style is wrong, regenerate or restyle the whole image instead. Inpainting a picture that is mostly wrong takes longer than starting again.

List the fixes and do them one at a time, biggest first. One mask, one change, per pass. Fixing a hand and a sign in one mask gives the model two jobs and a muddled prompt.

Paint the mask a little bigger than the problem

Cover the whole problem plus a margin of clean pixels around it, so the model has room to join old and new. For a hand, include the wrist. For an object you are removing, include its shadow and reflection, or they stay behind as ghosts.

Don't cover a face or a product you want to keep. Everything inside the mask can change.

Soften the edge

A hard mask edge leaves a visible seam where new pixels meet old. A soft edge blends them. Firefly's brush has a Brush hardness slider: lower it. Automatic1111 and Forge have Mask blur (default 4, up to 64). ComfyUI's VAE Encode (for Inpainting) grows the mask with grow_mask_by (default 6, up to 64).

Tools without an edge setting (Midjourney, ChatGPT, Magnific, Krea, Ideogram) still blend better when the mask runs a few pixels into clean background instead of hugging the object's outline.

Describe what should be there, not what is wrong

The model reads your prompt as a description of the masked area. "Fix the hand" tells it nothing about what a good hand looks like. "A relaxed left hand resting on the table, five fingers, soft window light from the left" does. Mention the light and the material so the new part matches the old.

To remove something, describe the background that should show through ("empty wooden floor"), or in Photoshop and Magnific's eraser leave the prompt blank and the tool fills from the surroundings.

Here is my image and what's wrong with it: [DESCRIBE]. Plan the inpainting passes in order, one change per pass. For each pass tell me exactly what to cover with the mask (including any shadow, reflection or wrist), and write the fill prompt describing what should be in that area: the object, its pose, the material and the light direction you can see in the rest of the picture. I'm using [TOOL].

Before

After

Set the strength, where the tool lets you

In Automatic1111 and Forge, Denoising strength (default 0.75) sets how far the new pixels may move from the old. As a starting point we use about 0.4 to 0.6 with Masked content set to original for a small correction, and 0.8 to 1 with fill or latent noise to replace an object outright. Inpaint area "Only masked" (padding default 32 px) redraws the masked part at full resolution, which helps small details such as faces in a wide shot.

In ComfyUI, keep denoise high with VAE Encode (for Inpainting): the docs warn that a low value leaves grey blocks. For gentle fixes, encode normally and use Set Latent Noise Mask or InpaintModelConditioning instead. On the FLUX.1 Fill API, steps run 15 to 50 (default 50) and guidance 1.5 to 100 (default 60).

Photoshop, Firefly, Midjourney, ChatGPT and the other app tools don't expose a strength setting. Control them with the mask and the prompt.

Generate several and check at 100%

Ask for three or four variations and zoom to 100% before choosing. Check the edge for a seam, the light direction, the colour temperature, the grain or noise, and the contact shadow where the new part touches something.

If one is almost right, don't regenerate the whole area. Mask just the remaining fault, smaller, and run another pass. Upscale only after the fixes are done.

Here are the original and my inpainted version. Compare them closely: is there a visible seam at the mask edge, a change in light direction, colour temperature, sharpness or grain, a missing contact shadow, or anything that changed outside the area I meant to edit? List each problem with where it is, and tell me the smallest mask and prompt that would fix it.

Where inpainting lives in each tool (as of October 2026)

Photoshop (desktop): make a selection, for example with the Selection Brush, select Generative Fill, pick a model in the Contextual Task Bar (Firefly Fill & Expand, Firefly Image 5 or 1, or partner models FLUX.2 pro, FLUX Kontext pro, Gemini 3.1 with Nano Banana 2, Gemini 3 with Nano Banana Pro), type a prompt or leave it blank to fill from the surroundings, then choose a variation in the Properties panel.

Adobe Firefly (web): Image > Edit image, then Fill in the left toolbar. Brush over the area (Select more / Select less), set Brush size and Brush hardness, optionally add a reference image, type the prompt and select Fill. Removal has its own Generative remove tool. Hover the indicator by the Fill button to see the credit cost.

Midjourney: in the Editor on midjourney.com (the Edit tab for your own uploads), use the Erase brush (Restore brings areas back) or Smart Select with include and exclude points, then Erase Selection. Edit the prompt and submit. Inpainting uses the V8.X Edit Model; V8.X edits appear on your Create page and publicly unless you use Stealth mode.

ChatGPT Images: open the image, choose Select, highlight the area (a slider sets the size), then describe the change. OpenAI warns that highlights are not always precise and edits may extend beyond the selection, so check the rest of the picture.

Gemini (Nano Banana): Google's help documents no mask tool. Edits are made by description ("Local edits"), so name the area precisely: "the mug on the left, only the mug". Editing images requires being 18 or over.

Magnific (formerly Freepik, renamed in April 2026): Image Editor > Retouch. Mask with Lasso, Click or Brush, describe the change, Generate. Include Image puts a reference photo into the mask; Character mode uses your Custom Characters. To remove something, use Edit's eraser: mask and Generate with no prompt.

Krea: Krea Edit > Change Region. Select the whole image, paint with the brush (adjustable size), draw a rectangle, or hover for Auto Mask, then prompt and Generate. Earlier results stay as thumbnails to compare or revert.

Ideogram: Studio (it replaced Canvas and Magic Fill). Select the image layer, then Tools > Select area (paid plans), set Brush size (10 to 80), paint, describe the change and select Generate. Only the painted area changes.

FLUX.1 Fill [pro] (Black Forest Labs API, /v1/flux-pro-1.0-fill): send the image and a black-and-white mask of the same size, white is redrawn and black is kept, or one PNG or WebP where transparent areas are redrawn. steps 15 to 50 (default 50), guidance 1.5 to 100 (default 60), seed optional.

Automatic1111 and Forge: img2img > Inpaint. Mask blur 4 (0 to 64), Mask mode Inpaint masked, Masked content original (or fill, latent noise, latent nothing), Inpaint area Whole picture or Only masked, Only masked padding 32 px (0 to 256), Denoising strength 0.75. Soft inpainting has been built in since version 1.8.0. Upload mask expects white where you want inpainting.

ComfyUI: right-click Load Image > Open in MaskEditor, paint, Save. Feed the mask to VAE Encode (for Inpainting) (grow_mask_by default 6) with denoise kept high, or for subtle fixes to Set Latent Noise Mask or InpaintModelConditioning (noise_mask on by default). The official tutorial uses a dedicated inpainting checkpoint for cleaner transitions.

When it goes wrong

Visible seam or halo: the edge is too hard or too tight. Soften it (blur, hardness, grow) and repaint a thin mask across the seam only.

New part is too bright, too sharp or the wrong colour: describe the light and the camera in the prompt ("dim tungsten light, slight film grain"). In Automatic1111 and Forge, Only masked can over-sharpen a small area; try Whole picture.

The object comes back: the prompt still mentions it, or the shadow is outside the mask. Describe only the background, and include the shadow.

Hands or faces still melt: mask bigger (wrist, neck), describe the pose plainly, and generate more variations. Small faces in wide shots redraw better with Only masked, at full resolution.

Nothing changes: in Automatic1111 and Forge, denoising is too low or Masked content is original on a big change. Raise it or switch to fill or latent noise.