Editing, Consistency and Upscaling

The first image a generator returns is almost never the one that gets used. Usually it is 90 percent right with one ruined detail, or right in content and wrong in shape, or right on its own and wrong next to the other four images in the series. Each of those has a different fix.

Regenerate, inpaint or outpaint

Inpainting

Inpainting means painting a mask over the part to change, typing what should be there, and letting the model regenerate only the masked pixels. It blends edges with the surrounding image. Adobe's Generative Fill in Photoshop works this way (select an area, type a prompt), and OpenAI's image edits endpoint can modify an image partially or entirely from a new prompt. Chat tools like ChatGPT and Gemini also accept "change only the sign on the left to read CLOSED" in plain words, which works surprisingly often and fails surprisingly often.

How it behaves:

• A mask drawn tight to the flaw tends to leave a seam. A slightly wider one blends better.
• The prompt describes only what belongs in the masked area, not the whole image: "A relaxed hand resting on the table, five fingers."
• Results vary run to run, and inpainting is cheap, so users typically run it three or four times and pick.

Outpainting

Outpainting grows the canvas. A square becomes a wide frame, and the model invents what lies beyond the old edges. Adobe calls its version Generative Expand. It is the right move when a good 1024x1024 image needs a 1920x1080 version.

Keeping a character or style consistent

This is the hard part. Ask for "a tired woman in a green jacket" five times and you get five women. An explainer series, a reel with several stills or a recurring mascot needs the same one every time.

Tools offer a few handles for this, and they work best together.

1. Exact repeated wording

One fixed character description, pasted word for word into every prompt, gives the model the same target each time.

Character: Rosa, a woman about 60, short gray hair, round wire glasses, olive green canvas jacket, red scarf.

Distinct, nameable features carry over best: a red scarf, round glasses, a specific hat. "Kind face" carries over not at all.

2. Reference images

A reference image is a picture handed to the model as a visual target. Most leading tools in 2026 accept them. Google's Veo 3.1 documentation lists up to 3 reference images to preserve a subject's look in video. FLUX.2 is known for taking several references at once. ChatGPT and Gemini accept uploaded images in the chat and can be told "keep this character exactly, change the setting to a hardware store."

The usual approach is to make the first good image of the character the reference for every later one. A plain, front-facing, evenly lit image works best as a reference. A dramatic, half-shadowed image teaches the model the shadows too.

3. Fixed seed and settings

Where a tool exposes a seed, locking it while only the setting text changes helps. It will not guarantee the same face, and Midjourney's docs warn seeds can drift between versions and modes. It narrows the spread.

STYLE: flat vector illustration, clean shapes, subtle paper grain
PALETTE: slate blue, ochre, off-white, charcoal accents
LIGHT: soft overcast, no dramatic shadows
CHARACTER: Rosa, a woman about 60, short gray hair,
  round wire glasses, olive green canvas jacket, red scarf
FRAME: 1920x1080, Rosa on the left third, open space right
REFERENCE: rosa-front-v3.png (attach every time)
NO: text, logos, real place names

Checking a series side by side

Drift shows up when the finals sit side by side in a row. The usual differences:

• Same jacket color, same scarf, same glasses shape
• Same age (models drift younger, often)
• Same palette, same light direction
• No surprise text, logos or real landmarks

Inpainting usually fixes drift (the jacket goes back to olive) faster than regenerating from scratch.

Upscaling and what it cannot do

Upscaling enlarges an image. Classic resizing stretches pixels and gets blurry. AI upscalers instead predict what detail "should" be there: sharper edges, skin texture, brick lines. For an illustration, that is fine. It polishes something that was invented anyway.

For anything presented as real, it is poison. The upscaler does not recover detail that the camera missed. It invents plausible detail. A blurry face becomes a sharp face that may not be the person's face. A smudged plate number becomes a crisp number that may be wrong.