ICC Profile: Stop Color Drift in AI Image Exports
The orange jacket looked right in the editor, but the exported image looked dull in a browser. Was the file changed, the display different, or the ICC profile missing? Those are three different diagnoses. Before regenerating the artwork, inspect the downloaded pixels and their color description, then test a controlled copy in the receiving channel.
Create a source image in APOB, then check its downloaded color profile.
This guide is a test procedure, not a report that a particular APOB export failed. A publishable case study would need the source download, a metadata readout, two controlled derivatives, browser observations, and a print proof. The sample log below is deliberately labeled illustrative until those files exist.
Inspect the profile attached to the source image.
An ICC profile tells a color-managed application how to interpret image values. Two files may contain the same numerical RGB values yet be displayed differently when their color descriptions differ. That is why “it is red on my monitor” is not a sufficient asset specification. Adobe's documentation on embedded profiles and imported images gives a reliable starting point for distinguishing tagged, untagged, and converted files.
embedded ICC
Download the original from the APOB AI Image Generator, the portrait workflow, or the asset's actual source route. Preserve that binary before opening it in an editor that might auto-convert it. Record filename, dimensions, format, checksum, and whether an embedded profile is present. If present, record its exact name, not “RGB.” If the inspection tool says “sRGB IEC61966-2.1,” copy the complete label. If it says none, write untagged; do not infer sRGB merely because the image looks normal on one screen.
This check is pending. APOB is not claimed to expose an ICC selector. Inspect any upscaled derivative separately.
untagged file
An untagged file lacks an embedded statement of what its pixel values mean. A receiving application must guess or apply a default working space. A correct-looking image on the creator's machine may therefore diverge elsewhere. Adobe's imported-profile guidance treats untagged documents separately from tagged ones. When opening a file, note whether the editor prompts to assign a profile, convert to its working space, or leave values untouched.
Do not describe a guessed assignment as recovered color. Assignment changes interpretation; conversion recalculates values. Preserve an ambiguous original and document the decision.
display profile.
The image's ICC profile is not the same as the monitor's profile. A calibrated display profile characterizes the device used for preview; the file profile describes the image data. Record the display model, profile/date if known, operating system, application, and whether color management is enabled. A screenshot from one uncalibrated laptop is weak evidence of how a print or another device should look.
When reviewers disagree, compare applications on one display before blaming the source. Web art needs a browser check; print art needs a supplier proof.
Reproduce one color shift with controlled copies.
The controlled experiment changes one variable at a time. Make a tagged working copy, duplicate it, remove the profile from the duplicate without changing pixel values, and open both in the same color-managed application. Keep the master untouched. Do not publish a “before/after” pair without recording the actual file operations; otherwise an export's compression, resampling, or appearance adjustment could be mistaken for an ICC effect.
same pixels
Start from one downloaded image and duplicate it at the file level. Record a source checksum, then use a tool that can show both pixel dimensions and profile state. For a rigorous same-pixels comparison, confirm the decoded RGB raster values are identical before changing metadata; byte-level file hashes will naturally differ when metadata changes. Fix the viewing application, display, zoom, and background. If the tool silently reconverts or recompresses on save, the pair is no longer an isolated profile test.
A useful test asset contains skin, a saturated color, neutral gray, and a permitted brand swatch. The pair demonstrates interpretation, not objective color accuracy.
profile removed
On a copy only, remove the embedded profile while leaving the raster values unchanged. Label it untagged_test, not web_ready. Then open the tagged and untagged copies in the same color-managed viewer and one target browser. Note if the application asks how to handle missing profiles. Preserve screenshots of its choice dialog and displayed files. If no visible shift appears, report “no difference observed in this environment” rather than claiming untagged files are universally safe.
A profile-stripped copy is diagnostic, not the delivery file. See Adobe's embedding instructions.
annotated difference.
Create an annotation sheet with fixed sample points: face midtone, saturated jacket, white background, and brand swatch. For each, record the two filenames, profile state, viewer, screen/profile, observed direction of change, and reviewer. Use a neutral caption such as sample 03 | jacket orange | tagged appears warmer in viewer X | untagged appears less saturated; do not invent numeric color-difference scores without a measured instrument and defined color space.
No annotated APOB pair exists yet. An illustrative record is FAIL — source untagged; browser differs; cause pending. One improved screen is not proof of repair.
Convert a web version while retaining a print master.
One image may have two destinations with different requirements. Keep a tagged master and branch derivatives from it. Adobe's current online color workflow recommends sRGB for broad web compatibility. A print provider may specify a different profile, file format, and proof process. Neither destination should overwrite the master.
sRGB conversion
If the source has a known, embedded profile, convert a copy to sRGB with a color-managed editor for web delivery. Select convert, not merely assign, when the aim is to maintain the source appearance within the target space. Embed the sRGB profile in the exported file where the chosen format and workflow permit it. Record the editor version, rendering intent, export format, and any compression. Reopen the exported binary and confirm its actual profile; a checked box in the export dialog is not final-file evidence.
This draft does not claim APOB outputs Display P3, sRGB, or any particular space. An untagged file needs a documented interpretation before conversion.
print-provider profile
Ask the selected printer whether they want a tagged RGB file, a specific supplied profile, or a proofing conversion performed by their prepress team. Do not download an arbitrary ICC file and assume it represents the paper, ink, and press combination. If a provider supplies a profile, store its version and written instructions; soft proof a copy under their conditions if your editor supports it. Confirm whether they expect you to convert or leave the file in a known RGB space.
Out-of-gamut color, paper white, and lighting can change perception. Critical print work needs a physical proof.
export naming.
Name files by source, destination, and revision rather than by subjective appearance: asset-042_master_tagged.tif, asset-042_web_srgb_v2.jpg, asset-042_print_providerA_v1.tif. These are templates, not files generated for this article. Keep a manifest with pixel dimensions, embedded profile, conversion step, checksum, reviewer, and approval date. An export that is renamed without being inspected is not a verified web derivative.
Verify the receiving channel, not only the editor.
The last check happens where the audience or printer will actually consume the file. That check can reveal a mistake in export settings, profile stripping, a CMS image transformation, or an application that handles untagged media differently. Do not promise identical colors on every phone or uncalibrated monitor. Instead, record which environments were tested and which remain unknown.
browser preview
Upload the web derivative to a non-public preview where possible and open the delivered URL, not merely the local file. Compare the browser-rendered image with the approved tagged reference on the same calibrated display. Record browser version, operating system, display profile, asset URL, and whether the server transformed the image. A CMS thumbnail and the full-resolution asset may follow different conversion paths; inspect the one that will appear on the page.
If a preview differs, retrieve the served file and inspect its dimensions and profile. This separates “the browser displayed a tagged file differently” from “the delivery pipeline altered the file.” Do not call it an APOB fault without tracing the handoff. The APOB image workflow supplies source art; the publishing pipeline may be a different system.
recipient application
Open the same exported file in the recipient's actual application: the target social scheduler, marketplace upload preview, or printer intake portal. Capture the displayed version and any warnings. If the channel strips profiles or recompresses media, treat that as a delivery constraint and design a documented derivative for it. Confirm the channel's own specifications rather than assuming a generic sRGB export solves all color drift.
The reviewer must know which file and channel copy they are approving. Track crop and format changes separately from ICC issues.
proof checklist.
For print, request the provider's approved proof on the selected stock. Inspect skin, saturated garment, shadows, neutral backgrounds, and any brand swatch under agreed lighting. Mark each item pass, revise, or reject. Keep the proof date and supplier contact. If an issue is really a paper or lighting effect, changing a source ICC profile blindly may make the web asset worse.
Sign-offs should name derivative hashes. Re-exporting invalidates the previous check.
Preserve a color provenance record for revisions.
The useful outcome is a repeatable decision trail. A designer joining the project next month should know where the image came from, how its colors were interpreted, what was converted, what each recipient reviewed, and why a revision happened. A short log is sufficient if every field points to a real file or screenshot.
source checksum
Hash the source and derivatives; save permission, date, route, dimensions, and format. Redact private URLs in the public article.
profile ID
Use exact profile names and label the operation: embedded profile inspected, profile assigned, or converted to destination profile. Those are materially different actions. An illustrative row could be v1 | source untagged | interpreted as sRGB for diagnostic copy only | no approval; a repaired row requires a documented source interpretation and receiving-channel check. These are example rows, not results from an APOB export test.
approval date.
Record reviewer, date, channel, and evidence link. The final row might read web v3 | browser preview URL | approved on calibrated display | reviewer/date; leave it blank until the real preview occurs. For print, attach the printer's specification and proof approval. The governing rule is simple: no profile assumption becomes a fact through repetition. Create the image in APOB if it fits the brief, then verify the downloaded file and each destination independently.
