Most color grading examples compare unrelated pictures, so the viewer cannot tell whether mood came from the grade, the lighting, the subject, or the crop. A useful comparison begins with one neutral image and four written look contracts. The source, dimensions, protected colors, and review questions stay fixed. Only the requested treatment changes.

Build your own four-look proof sheet with APOB

This guide is a reproducible worksheet, not a claim that one prompt will behave identically on every asset. It shows how to prepare a color grading before and after review, preserve the rejects, and move any calibrated matching or LUT work into a finishing tool. Evidence and product pages were checked on September 11, 2026.

Choose a neutral anchor worth protecting

A neutral anchor is not necessarily flat or dull. It is simply a source with enough visible information to expose both success and damage. Use an image you own or have permission to edit. Keep the original file beside every generated version, and avoid comparing a low-resolution screenshot with a clean export.

Source criteria

Pick one frame that contains four kinds of evidence: a face or skin-like reference, a recognizable product color, a bright area with recoverable detail, and a shadow with texture. A plain studio portrait beside a colored package works better than a foggy silhouette because the reviewer can see what moved.

Record the filename, dimensions, crop, color profile if known, and capture date. Do not silently resize between looks. If the starting image already has crushed shadows or clipped highlights, flag that limitation before generation instead of blaming the later treatment.

Protected swatches

Choose two or three small regions that are not allowed to wander: a cheek, the product label, and a neutral gray or white object. Describe them in plain language and save close crops. The goal is not laboratory measurement; it is a visible contract that stops “more cinematic” from becoming an excuse for orange skin or a changed brand hue.

The APOB AI Color Grading page describes image and video editing paths, but it does not turn the workflow into a calibrated color suite. Treat the protected swatches as editorial review aids, not instrument readings.

Neutral reference

Place the untouched anchor in the first cell of the proof sheet and label it “source.” Use the same display size as the four variations. Reviewers should be able to switch their eyes between the source and a candidate without zooming, scrolling, or remembering an image from another page.

Keep a second copy with no arrows or annotations. That clean reference is useful when an annotation covers exactly the region under discussion. If the source contains text or a logo, add it to the protected-detail list even when color—not typography—is the main purpose of the test.

Reject conditions

Write automatic rejects before seeing output. Examples: face identity changes, product hue crosses into another brand color, legible text mutates, highlights lose necessary detail, or a shadow becomes a featureless block. A striking look that violates one of those rules does not advance to mood scoring.

Add a stop rule as well: if two repair attempts damage a different protected detail, return to the source or narrow the brief. Endless prompt polishing creates a misleading survivor set and hides the cost of failed outputs.

Write four moods as visual contracts

Mood names are too loose on their own. Convert each into observable intent for temperature, contrast, saturation, highlights, shadows, and protected details. These color mood examples deliberately occupy different parts of the creative space while sharing the same source.

Clean commercial

Contract: neutral-to-slightly-warm whites, restrained contrast, moderate saturation, open shadows, controlled highlights, and faithful product color. The image should feel ready for a clear catalog or campaign layout, not dramatically stylized.

Prompt template: “Create a clean commercial grade with balanced neutral whites, natural skin, open shadow detail, controlled highlights, and unchanged product color. Preserve face, label, typography, crop, and object geometry.” Use the words as direction, then judge the rendered evidence rather than assuming each adjective has a fixed numeric meaning.

Warm lifestyle

Contract: a gentle warm bias in ambient light, softer contrast, slightly richer midtones, and comfortable skin without yellowing neutral objects. Warmth should support intimacy, not repaint the product.

Prompt template: “Apply a warm lifestyle treatment with soft golden ambience, relaxed contrast, rich but believable midtones, and protected skin and product color. Keep whites readable and preserve the original face, label, crop, and geometry.” If the result warms everything equally, repair the ambient description rather than asking for “more warmth.”

Cool nocturne

Contract: cool environmental shadows, a readable subject, contained highlights, and enough separation that the product does not disappear. The face may sit in a cooler scene, but it should not turn gray or cyan.

Prompt template: “Create a cool nocturne mood with blue-leaning environmental shadows, a neutral readable subject, contained practical highlights, and visible texture in dark areas. Preserve skin, product hue, label, crop, and geometry.” This is a mood treatment, not proof of a calibrated cinematic color grade.

Editorial contrast

Contract: stronger tonal separation, decisive blacks, selective saturation, and crisp subject emphasis without clipping important detail. It should feel graphic at first glance and still survive close inspection.

Prompt template: “Apply high-contrast editorial color with firm black points, controlled bright detail, selective saturation, and clear subject separation. Do not alter skin identity, brand hue, text, crop, or object shape.” Keep the clause about detail; “high contrast” alone can reward a dramatic but unusable result.

Generate the proof sheet from one locked source

Run the four contracts through the same verified edit path. An AI Image Generator can create the source when the rights and identity are clear, but once the test begins, freeze that source. Do not swap in the most flattering input for each mood.

Keep-original setup

Upload the approved anchor through the displayed keep-original or image-edit workflow. Record the page used, model or mode label, aspect ratio, visible strength control if any, displayed credit estimate, generation date, and prompt. If the interface does not expose a seed, write “seed unavailable” rather than pretending the runs are deterministic.

Name outputs before judging them: source, clean-v1, warm-v1, cool-v1, and editorial-v1. The filename keeps preference from rewriting the history later. Store the original download, not a compressed preview screenshot.

One-pass outputs

Generate one first pass per contract. Do not cherry-pick from a large hidden batch. Place all four first attempts in the evidence folder, including obvious failures, and note the displayed cost or credit count exactly as shown. If pricing or credit display varies by account or region, record the observation without generalizing it.

Create a quick ledger with columns for prompt version, preserved details, reject reason, and reviewer status. The first pass establishes how much repair the brief needs; it is not a product-wide benchmark.

Second-pass repairs

Repair only one failure at a time. If the cool look changes the package, add a product-hue clause while leaving every other sentence untouched. If the warm look clips highlights, revise only the highlight instruction. A one-change retest makes the before/after relationship readable.

Keep v1 beside v2. Never overwrite the rejected file. A repair is successful only when the target problem improves without introducing another automatic reject. When a new defect appears, mark the version “tradeoff,” not “fixed.”

Contact-sheet assembly

Build a two-row sheet: source plus first passes on top, accepted repairs or explicit rejects below. Give each cell the same size and include a short code, not a persuasive label. Put the full prompts and settings in a separate table so the pictures remain large enough to inspect.

For moving assets, save stills at matched moments before opening the APOB Colour Correction workflow. A frame sheet can reveal color drift, but it cannot prove that the look stays stable through motion.

Blind-review mood and brand safety

Remove mood names from the sheet and randomize the four candidate positions. Reviewers should know the intended definitions but not which prompt produced which cell. That small separation limits the tendency to reward an output merely because its label sounds right.

Mood match

Ask, “How well does this image satisfy the stated visual contract?” Score 1 for contradiction, 3 for partial fit, and 5 for a clear fit with no explanation needed. Add one sentence of evidence: “Shadows are cool, but the subject reads neutral,” for example.

Do not average a tiny group into fake precision. Show individual scores and the number of reviewers. If only one editor completed the sheet, call it an editorial review, not consensus.

Skin check

Compare the face at full image size and at a close crop. Check hue, local contrast, texture, boundary artifacts, and identity. A mood may alter surrounding light while still preserving a believable person; it fails when the grade makes the face look like a different asset.

Review this dimension before aesthetic preference. A beautiful frame with unstable skin is not a safe reusable recipe for creator-led content.

Product-color check

Place the source swatch and candidate swatch side by side. Ask whether a customer could mistake the candidate for a different product variant or brand color. Include text legibility and logo shape in the same check when those elements appear.

This is a visual editorial gate. For contractual color matching, print work, broadcast delivery, or shot-to-shot normalization, move the selected look into a managed color workflow rather than treating an AI preview as the final authority.

Mobile thumbnail check

Reduce every candidate to the approximate size used in a mobile feed. Check whether the subject remains visible, dark regions collapse, highlights dominate, or product color loses separation. A look can pass at desktop size and fail as a thumbnail.

Record both decisions. “Desktop pass, mobile fail” is more useful than a single total score because it tells the team whether to adjust the grade, crop, or destination.

Turn the winning look into a repeatable recipe

The deliverable is not the prettiest cell. It is a versioned instruction that another editor can apply to an authorized source and audit against the same protected details.

Reusable prompt

Copy the accepted prompt exactly, then attach the source conditions that made it useful: portrait plus product, neutral reference, visible highlights, and textured shadows. Add the displayed mode, settings, date, accepted output filename, and destination.

Call it look-recipe-v1, not “final forever.” AI color grading behavior and available models can change. A version label turns drift into something a team can detect.

Repair clause

Append the smallest successful repair as a conditional clause: “If product red drifts toward orange, reinforce unchanged product hue and label.” Do not merge every rejected instruction into one oversized prompt. A recipe should preserve the reason each clause exists.

Include the rejection evidence path beside the clause. That lets a future reviewer remove a workaround if the underlying behavior changes.

Image-versus-video note

State whether the recipe was tested on one still, matched video frames, or a complete moving clip. A still-image pass does not establish temporal stability. For video, review several moments and listen for any audio change if the chosen workflow touches audio.

Keep edit decisions separate from generation claims. A conventional dissolve or shot match can solve a sequence problem that prompt wording cannot.

Finishing boundary

Use the proof sheet to select direction. Use a professional finishing environment when the job requires calibrated monitoring, scopes, color management, LUT delivery, legal range checks, or shot matching. Blackmagic’s official DaVinci Resolve training explicitly covers scopes, balancing, color management, and delivery; its Colorist Guide provides exercise-led practice.

The boundary is a strength, not a disclaimer. APOB helps a creator explore and preserve an approved visual direction; finishing tools verify technical delivery. Save the source, four contracts, all outputs, blind scores, accepted recipe, repair clause, and finishing note together. That packet is the real value of these color grading examples.

Sources

• APOB AI — AI Color Grading

• Blackmagic Design — DaVinci Resolve Training

• Blackmagic Design — The Colorist Guide to DaVinci Resolve 20