Fix the Hand, Keep the Picture: How Inpainting Strength Works

You finally get the image. The light is right, the composition is right, the expression is right, and the subject has six fingers. The instinct is to press generate again. That is the most expensive fix available: a new seed changes everything, and the odds of getting the light, the composition and the expression back are worse than the odds of the hand you started with.

Inpainting fixes the hand and keeps the picture. You paint a mask over the part that is wrong, and the model regenerates only that part. The quality of the fix comes down to two decisions: where the mask goes, and how much you let the model forget.

What Is Inpainting?

Inpainting is image-to-image generation restricted by a mask. The Diffusers inpainting guide gives the convention: "the area to inpaint is represented by white pixels and the area to keep is represented by black pixels." Outpainting is the same idea pointed outward: extend the canvas, mask the new empty area, and let the model continue the scene.

The idea underneath both comes from SDEdit (Meng et al., 2021), which "first adds noise to the input, then subsequently denoises the resulting image" to make it more realistic. The authors name the central problem plainly: "balancing faithfulness to the user input" and "realism of the synthesized image." The more noise you add, "the synthesized images are more realistic but less faithful."

What Does Strength Actually Mean?

Strength is that noise dial. The Diffusers guide defines it as "a measure of how much noise is added to the base image, which influences how similar the output is to the base image." It also sets how much of the step schedule actually runs. The image-to-image guide gives the arithmetic: with 50 steps and a strength of 0.8, the pipeline adds "40 (50 * 0.8) steps of noise to the initial image and then denoising for 40 steps."

At the top of the range, the inpaint pipeline reference is blunt: at strength 1, noise is at maximum and "A value of 1 essentially ignores image." The masked area is generated from scratch, guided only by the prompt and the pixels around it.

The step counts are our own arithmetic at the 25-step default; the behaviour column follows the SDEdit trade-off. The paper's own example lands in the middle, suggesting a setting between 0.3 and 0.6 for its stroke-to-image task. The guide also warns about the low end: less noise means "the image quality may not be as great and the generated image resembles the base image more". Too low and the hand keeps its sixth finger.

The Mask Edge Is Where Fixes Show

A fix rarely fails in the middle of the mask. It fails at the border, where new pixels meet old ones. Three habits make the seam disappear:

1. Mask a little wider than the problem. Give the model some of the surrounding context to blend into, not a hard outline of the finger.
2. Soften the edge. The Diffusers guide notes that increasing mask blur softens "the transition between the original image and inpaint area", while zero blur "preserves the sharper edges of the mask."
3. Work small areas at higher resolution. The guide's padding_mask_crop option crops the masked area, upscales it for inpainting and overlays it back, which "is useful when the masked area is small while the image is large".

Regular Checkpoints Versus Inpainting Models

Any checkpoint can inpaint, but some are trained for it. The Stable Diffusion inpainting model card explains the difference: its UNet "has 5 additional input channels (4 for the encoded masked-image and 1 for the mask itself)", so the model sees the mask directly instead of inferring it.

The Diffusers guide is honest about both sides. With a regular checkpoint you "can still see the outline of the area the model is supposed to inpaint." Dedicated checkpoints create "a more natural transition", and as a result "are more likely to change your unmasked area." There is no free lunch, only a choice of which side of the seam pays.

For FLUX.1, Black Forest Labs ships a dedicated model. FLUX.1 Fill [dev] is "a 12 billion parameter rectified flow transformer capable of filling areas in existing images based on a text description." Its card lists its own limits: "slight-color shifts in areas that are not filled in", and complex textures that "may produce lines at the edges of the filled-area." Its weights carry the FLUX.1 [dev] Non-Commercial License, so read those terms before commercial work.

How Inpainting Works in Vision Studio

In Vision Studio, inpaint and outpaint use the checkpoint you have selected, and FLUX inpainting switches to FLUX.1 Fill automatically. Strength runs from 0.05 to 1.0, and the engine's default for edits is 0.75, the "replace it" end of the table, so lower it when you want to repair rather than replace.

Masks are layers on the canvas, alongside ControlNet guides and reference images, and the next render reads them. That means a fix can keep the composition guide that made the original work. Generative fill, object removal, AI expand and background replace are built on the same guided passes through your selected checkpoint.

A practical loop: lock the seed, mask the hand a little wide, start at 0.5, and step up by 0.1 until the fingers are right. Each change is one variable, so you know which one fixed it.

Repair, Do Not Re-Roll

A re-roll is a lottery ticket against everything you already won. Inpainting spends the effort only where the image is wrong. Mask generously, start in the middle of the strength range, and treat 1.0 as the tool for filling space, not fixing detail.

Vision Studio is free, open-source (MIT), and runs locally on your GPU. Inpaint, outpaint and the canvas control layers are documented with their source files at vision-studio-x.com/features. Download it at vision-studio-x.com/download.

-Rocky

#Inpainting #FLUX #StableDiffusion #ImageEditing #LocalAI #OpenSource #VisionStudio