Level 6Lesson 50⏱️ 45 min

Advanced Image Workflows

Editing, inpainting, reference images, upscaling, and producing brand-consistent assets at scale

Beyond Generation: Editing What Exists

Generating a fresh image is the easy 20%. The professional 80% is editing: fixing the one wrong hand, swapping a background, keeping a character consistent across twelve images, and matching a brand. This lesson is the toolkit for that work.

Key idea: Modern image tools are increasingly editors, not just generators. The difference between an amateur and a pro is almost entirely in the edit pass.

Inpainting & Outpainting

Inpainting regenerates a masked region while leaving the rest untouched - the surgical tool of image editing. Outpainting extends the canvas beyond its original borders, inventing what lies outside the frame.

INPAINTING - fix or change part of an image
  1. Mask the region (the awkward hand, the logo, the sky)
  2. Prompt ONLY for that region: "a clean blue sky"
  3. Generate → only the masked area changes

OUTPAINTING - expand the frame
  Original is 1:1 → extend left/right to make it 16:9
  The model invents plausible surroundings
  Great for reformatting one asset to many aspect ratios
Where to do it: Photoshop's Generative Fill, Firefly, and most web image tools expose inpaint/outpaint. It's the fastest fix for "90% perfect" images.

Reference Images: Steering With Pictures, Not Words

Words have limits. Reference images let you transfer a style, a composition, or a character's face directly. This is how you get consistency that prompting alone can't deliver.

Style reference   → "make new images that LOOK like this one"
                    (Midjourney --sref, Firefly style ref)

Character/face    → keep the same person across many images
                    (character reference / "consistent character")

Structure/pose    → match the composition or pose of a reference
                    (ControlNet edges/pose in local tools)

Composition       → use a rough sketch as the layout skeleton
Consistency unlock: A recurring character or product shot is impossible with text alone. Reference images (or a trained style) are the only reliable path.

Upscaling & Cleanup

Generated images are often 1024px - fine for web, too small for print or large displays. Upscalers increase resolution while adding plausible detail (not just stretching pixels).

Upscale for resolution

Take 1k → 4k for print, banners, or zoomed detail. Built into most tools; dedicated upscalers (Topaz, Magnific-style) go further.

Remove background

One-click cutouts for product shots, logos, and compositing onto new backgrounds.

Fix faces & hands

Use face-restore / inpaint passes on the classic diffusion failure points before you ship.

Brand-Consistent Assets at Scale

The real business value is producing many on-brand assets, not one pretty picture. Consistency comes from constraints you reuse every time.

A repeatable brand recipe:

1. Lock a STYLE reference image (your brand's look)
2. Keep a fixed palette + descriptor block:
   "flat illustration, [#hex] and [#hex] palette,
    rounded shapes, soft shadows, friendly"
3. Reuse the same seed family for visual cohesion
4. Generate each asset, then inpaint to fix details
5. Run the same upscale + crop preset on all outputs

→ 30 social tiles that clearly belong together
Commercial reminder: For brand work you intend to publish, prefer a commercially-safe generator (Firefly) or assets you have rights to. Lesson 54 covers the licensing rules in full.

Hands-On: Rescue and Reformat One Image

Hands-on (20 min): Generate a hero image with one obvious flaw (a messy background or a bad object). Then: (1) inpaint to fix the flaw, (2) remove or replace the background, (3) outpaint it from 1:1 to 16:9, and (4) upscale the result. You've now taken a single "almost" image and produced a clean, reformatted, print-ready asset - the everyday workflow of anyone using AI images professionally.
Lesson 50 Quick Reference
Inpainting

Mask a region and regenerate only that part - the surgical fix for almost-perfect images

Outpainting

Extend the canvas beyond its borders; reformat 1:1 → 16:9 or 9:16

Style reference

Feed an image so new outputs match its look - consistency words can't achieve

Character reference

Keep the same face/character across many images

Upscaling

Increase resolution while adding detail; needed for print and large displays

Background removal

One-click cutout for product shots, logos, and compositing

Brand recipe

Locked style ref + fixed palette/descriptors + seed family + consistent post-processing

← AI Image Generation Fundamentals
Unlocks in ~12 min of reading