How to Use Fusion Brain Image-to-Image and Inpainting

November 10, 2026

Text-to-image gets the attention, but the image-to-image and inpainting modes are where Fusion Brain becomes a working tool rather than a toy. You can take a phone photo of a product on a messy desk, repaint the background, and keep the label text intact.

If you are new to Fusion Brain, read the text-to-image basics first; this guide assumes you have generated a few images already. It walks through redraw strength, mask painting, and the sequencing mistakes that waste twenty minutes each time.

What is the difference between image-to-image and inpainting?

Image-to-image redraws the whole picture, guided by your prompt and a strength setting. Inpainting redraws only the region you paint over, leaving everything outside the mask untouched.

Use image-to-image when the composition is right and the style is wrong. Use inpainting when 95% is right and one object is wrong. Reaching for image-to-image to fix a single object is the single most common time sink here.

What strength value should you use in image-to-image?

Start at 0.55 and adjust by 0.1 in either direction.

  • 0.3-0.4 — near-identical to the source, subtle texture and lighting shifts only.
  • 0.5-0.6 — same layout and subjects, new style or palette. The useful default.
  • 0.7-0.8 — loose reinterpretation; expect anatomy and proportions to drift.
  • Above 0.85 — effectively text-to-image with the original as a faint suggestion. Stop here and just use text-to-image.

Remember the direction: higher numbers mean the model overrides the input more. Some builds label this “denoising strength” or “similarity” with the scale inverted, so check which way the slider runs before you drag it.

How do you paint an inpainting mask that actually works?

Three habits.

  1. Cover generously. A mask that hugs the object boundary leaves a halo of the old pixels, and the model blends old and new into mush. Paint 10-20% beyond the object.
  2. Match the lighting in your prompt. If the original has a hard light from the left, say so. Inpainting fills a hole; it does not relight the scene.
  3. Describe the replacement, not the removal. “a white ceramic mug” not “no cup”. Negative descriptions rarely work in a mask.

Brush size matters more than people expect. Use a large brush for coverage, a small one for the edge, and accept that fine detail around fingers and hair will need a second pass.

What order should you do edits in?

Work from big to small, and save a copy between every stage.

  1. Redraw the whole image once for style, if needed.
  2. Inpaint the background or large regions.
  3. Inpaint individual objects.
  4. Fix small artifacts last, with tight masks.

Doing step 3 before step 1 wastes work: the style pass repaints your carefully fixed object. Save each intermediate as a separate file — Fusion Brain does not keep a version history in the free web app.

Can you keep text legible through inpainting?

Partially. Text inside the mask will be destroyed and replaced with plausible-looking shapes. Text outside the mask survives if you keep the mask edges away from it.

For a product label, mask everything except the label. If the label itself needs to change, plan on setting type in a separate editor afterward — no current image model reliably reproduces specific words unless you prompt the exact string, and even then results are hit-or-miss.

Where does this workflow fall short?

Faces. Inpainting a face at 1024×1024 with a wide mask usually produces a different person, and matching an existing face is not supported. Upscale limits also mean small source images (under 512px) come out soft no matter what settings you use.

The mask editor is also rough compared with dedicated tools: no feathering control exposed, no mask save-and-reuse. For repeated product reshoots on dozens of images, a batch-capable tool with an API pays for itself within a week. Fusion Brain’s strength is single-image, on-demand fixes with zero setup.

Quick Summary

  • Image-to-image redraws everything guided by strength; inpainting changes only the masked region.
  • Strength 0.5-0.6 is the default for restyling; above 0.85 you might as well use text-to-image.
  • Paint masks 10-20% beyond the object to avoid halos, and describe the replacement object rather than its absence.
  • Order edits big to small and save between stages; the free web app keeps no version history.
  • Faces and exact text do not survive inpainting — plan a separate pass for those.