In Class 57 you directed new images into existence. This class starts from the opposite position: you already have a photograph, screenshot, or product shot, and it is almost right. The light is dim. The background is distracting. An old print is scratched. The temptation is to hand it to an AI tool and say "make it better." That instruction is too vague to be safe, because "better" can quietly change what the image claims.

Enhancement is improving an existing image while preserving the details that must remain true. Generation creates a new visual. Enhancement repairs and clarifies one that already exists.

The nine operations, defined plainly

Students mix these terms up, and tools mix them further. Keep them distinct:

The first five can usually be done conventionally, with deterministic tools. The last four cross into invention. That does not make them forbidden. It makes them subject to one question: is the altered area material to the claim the image makes?

The core distinction: improve the rendering, not the facts

Improving lighting, noise, and crop is different from changing the product, person, event, evidence, or claimed result.

A dim exposure is a rendering problem. A missing product button is a fact problem. Noise reduction clarifies what the camera saw. Regenerating a label rewrites what the product says. Straightening a horizon helps the viewer. Moving a person closer together invents a relationship.

Three examples make the boundary concrete:

1. A dim product photo that needs cleaner light. A seller photographs a ceramic mug on a kitchen table. It is underexposed, yellow from indoor bulbs, with a distracting dish rack behind it. Safe work: white-balance correction, exposure lift, noise reduction, background cleanup, gentle shadow recovery. Unsafe work: changing the glaze color, straightening a handmade wobble, sharpening a printed logo into different lettering, or replacing the background with a marble countertop that implies a different lifestyle without disclosure. The buyer must receive the color and shape shown.

2. An old family photograph that needs repair. A scanned 1970s print is faded, scratched, with a torn corner over plain background. Safe work: dust and scratch removal, fade and contrast restoration, tear reconstruction from surrounding background, careful sharpening. Unsafe work: letting a restoration tool redesign faces, invent smiles, change clothing patterns, or rebuild the missing corner as a detailed room that never existed. Repair clarifies memory. It does not author new memory.

3. A chart or screenshot that must not be "beautified" into different data. A research story includes a bar chart screenshot. It is slightly blurry and the colors are dull. The correct fix is to re-export the chart at higher resolution or rebuild it from the same numbers — never to run it through a beautifier that straightens bars, sharpens digits into different digits, or "cleans up" an awkward outlier. A visual may illustrate a claim. It must not upgrade a weak claim into a stronger one. When the pixels are the evidence, regeneration is falsification.

The decision ladder

Use the least inventive method that solves the problem:

1. CONVENTIONAL EDIT FIRST
   exposure / white balance / contrast / crop / straighten
   noise reduction / sharpening / dust removal
   ↓ if insufficient
2. LOCAL REPAIR
   masked inpainting of dust, scratches, small distractions
   tight masks, content drawn from immediate surroundings
   ↓ if the frame or scene is still unusable
3. GENERATIVE EXTENSION / REPLACEMENT
   outpainting, background replacement, large-area fill
   ONLY when the altered area is not material to the claim
   + record it in the edit log

A practical rule: if a viewer would change their decision — buy, believe, hire, date, invest — because of the altered area, that area is material. Treat it as off-limits for generative replacement, or remake the photograph honestly.

Upscaling deserves its own caution. Modern upscalers do not merely enlarge pixels; they synthesize texture. On faces, fabric, and printed text, compare the upscaled result against the original at 100 percent. If the logo gained serifs, the skin gained pores the camera never captured, or small text became confident gibberish, reject it or mask those regions out.

The edit plan: preserve, improve, do not invent

Before touching an image, write a short IMAGE-EDIT-PLAN.md. It takes ten minutes and prevents most enhancement failures:

# IMAGE-EDIT-PLAN.md — Ceramic mug, listing photo

## Must preserve (facts)
- Mug geometry, handle shape, glaze color (#7A9B8E verified vs. physical mug)
- Printed logo text and placement
- Visible condition: small chip on base (disclose, do not erase)

## May improve (rendering)
- White balance, exposure, shadow detail
- Background cleanup behind product
- Crop to 4:5 + 1:1 variants, copy space top-left

## Must not invent
- No new glaze texture, no reshaped handle
- No regenerated logo lettering
- No lifestyle background implying materials/context not shown
- No removal of the chip; note it in the listing

## Method ladder
1. Lightroom-style global corrections
2. Local mask cleanup only
3. No generative replacement (background stays neutral sweep)

## Review
- Owner: [name] — compare export vs. physical product side by side

Then execute and keep the receipts: source file untouched, working file with layers or history, before/after pair at the same crop, and a concise edit log (tool, date, operation, masked area, setting). A reviewer should be able to reconstruct what changed without guessing.

Exercises

1. Write an edit plan before you edit. Pick one real image: a product shot, an old photo, or a document screenshot. Fill in the three columns — must preserve, may improve, must not invent — in IMAGE-EDIT-PLAN.md. Keep each column to five bullets or fewer.

Finish line: a plan another person could follow without asking what matters.

2. Produce a before/after pair with a log. Apply only ladder steps 1 and 2. Export the before and after at identical crops. Write five to eight log lines: operation, tool, area, and why it was safe.

Finish line: a before/after pair where every change is explainable in one sentence.

Common failure: running a one-click "AI enhance" first and discovering the product color shifted or text changed. Recovery: return to the source file, redo global corrections manually, and restrict AI tools to tightly masked local repairs.

Verify fast: place source and export side by side at 100% on faces, labels, and small text. Every change must map to one log line. Publish gate: no export publishes until the owner initials the must-preserve column — especially charts/screenshots where pixels are evidence.

Check your understanding

• What separates retouching from generative alteration, and why does the distinction matter for trust?
• Why is a blurry chart fixed by re-exporting rather than beautifying?
• When is generative background replacement acceptable, and what makes an area "material to the claim"?
• What three columns does IMAGE-EDIT-PLAN.md require, and what does the edit log prove?

In the next lesson you will apply this discipline to the two highest-stakes enhancement jobs: product images that must sell truthfully, and profile portraits that must flatter without inventing a different person.