Stable Diffusion Inpainting Online: Fix Hands & Faces

Why inpainting beats another full redraw

Text-to-image can give you a strong frame—and still leave one broken hand, a warped eye, or a logo you never asked for. Regenerating the whole image often destroys the composition you already like. Stable Diffusion inpainting online solves a different problem: redraw only the masked region while the rest of the canvas stays put.

On stable diffussioncn (stablediffussion.cn), open the online generator, switch to Inpaint, upload the image, paint a mask, write what should appear inside that mask, and tune Denoising Strength. Zero GPU install. That is the browser-native path this guide is built around.

Inpaint vs img2img: pick the right tool

Job	Prefer	Why
Restyle the whole frame, keep silhouette	Img2img	Global Strength rewrites everywhere
Fix one hand, face, label, or prop	Inpaint	Mask limits where pixels change
Swap backdrop but keep the subject sharp	Inpaint the background (or low-Strength img2img)	Mask protects the subject
Rebuild a failed pose from scratch	Txt2img + new Seed	Local edit cannot invent a new pose cleanly

Rule of thumb: if the failure is local, use an inpaint mask. If the failure is global style, use img2img. For Steps, CFG, and Seed basics, start from the online parameters guide.

How Stable Diffusion inpainting works (browser version)

1. Upload a finished or nearly finished image.
2. Paint a mask over the region to regenerate (brush / erase / clear on this site’s canvas).
3. Prompt for the masked area—describe what should appear there, not a full scene rewrite unless you are replacing a large region.
4. Set Denoising Strength (on this site roughly 0.20–0.95).
5. Keep CFG and Steps in a sane band (SDXL everyday: CFG ≈ 5.5–7.5, Steps ≈ 24–32), then generate.
6. If the patch still fails, adjust mask size or Strength—one variable per try—or regenerate the base image after three failed local passes.

Official Stability docs cover model families and APIs; this article is an independent browser workflow for creators using stablediffussion.cn, not a substitute for vendor docs.

Mask craft: the underrated half of inpainting

Bad masks cause more failures than “weak prompts.”

• Include context: for a hand, mask wrist and a strip of sleeve; for a face, include hairline and neck edge.
• One problem per pass: fix the left hand, then the right; do not mask both hands and the logo at once.
• Feather mentally: hard postage-stamp masks leave seams; slightly oversized soft regions blend better.
• Describe the replacement, not the defect: prompt natural five-finger hand holding a ceramic cup, not fix the weird fingers.
• Negatives stay short: extra fingers, fused fingers, deformed hand, blurry face—see the site’s negative prompt guide for baseplates.

Denoising Strength for local edits

Strength still means “how hard the model forgets the masked pixels,” but the unmasked area is locked.

Strength	Behavior inside the mask	Typical use
0.30–0.45	Mild rewrite	Soft skin polish, small texture cleanup
0.45–0.60	Balanced	Face touch-up, label cleanup, prop swap with structure cues
0.60–0.75	Strong rewrite	Broken hands, heavy face rebuild
0.75–0.90	Near full regen in mask	Remove / replace objects; large backdrop patches

On stable diffussioncn, ladder Strength (e.g. 0.50 → 0.60 → 0.70) with the same Seed when you need comparable A/B frames. Jumping straight to 0.90 often invents a hand that ignores the arm angle.

Three online workflows you can reuse

A) Fix hands without killing the pose

1. Crop or generate so the bad hand is readable.
2. Mask hand + wrist; leave the torso unmasked.
3. Prompt: anatomically correct hand, five fingers, matching skin tone, same lighting as reference.
4. Strength 0.60–0.75, CFG 6–7, Steps 28–32.
5. If fingers still fuse, widen the mask slightly and raise Strength by ~0.05—or hide the hand with sleeves / framing and regenerate.

B) Face and eye cleanup

1. Mask face soft edges; avoid masking the entire hairstyle unless hair must change.
2. Prompt: same identity, clear eyes, natural skin texture, coherent lighting.
3. Strength 0.45–0.60 first—faces melt faster than hands when Strength is high.
4. Pair with a short negative: asymmetric eyes, plastic skin, extra pupils.

C) Remove or replace a background object

1. Mask the object with a little surrounding backdrop.
2. Prompt the desired fill (empty marble table, soft studio gray backdrop), not “delete bag.”
3. Strength 0.65–0.85 so the old silhouette does not ghost through.
4. For commerce SKUs, protect logos with a tight unmasked region; only mask what must change.

Recommended loop on stable diffussioncn

1. Lock a good txt2img or img2img base (Seed noted).
2. Open Inpaint on the generator.
3. Mask one failure class.
4. Write a local prompt + short negative.
5. Ladder Strength; lock Seed once a candidate looks promising.
6. Export; only then start a second mask for the next defect.

Browse the prompt library for subject language and the tutorials for structured practice. Inpainting is where “almost good” frames become shippable without gambling a new Seed.

Common failures and fixes

Symptom	Likely cause	First fix
Obvious seam / color break	Mask too tight or Strength too low	Widen mask; Strength +0.05–0.10
Hand still broken	Strength too low or mask misses wrist	Strength 0.65–0.75; include forearm
Face identity drifts	Strength too high / prompt reshapes identity	Strength −0.10; say same person
Object ghost remains	Strength too low for removal	0.75–0.85 + describe the fill
Whole mood of the frame shifts	You used img2img globally	Switch to inpaint mask
Patch ignores prompt	Prompt describes the full scene, not the hole	Rewrite for the masked region only

Takeaways

1. Inpainting is local surgery; img2img is a full restyle—choose by failure size.
2. Mask quality + Strength band matter more than stacking quality tags.
3. On stable diffussioncn, the browser Inpaint path (upload → mask → Strength → generate) turns Stable Diffusion inpainting online into a repeatable finish pass.

Open the Inpaint mode, fix one hand or face region with Strength around 0.65, and keep the rest of the frame untouched. When you can explain why a patch worked, you are using online Stable Diffusion the way production workflows do—not as endless rerolls.