AI Recoloring Product Photos Without Breaking Brand

The ecommerce team had forty SKU photos shot on a neutral gray sweep. The rebrand shifted primary from cobalt to a greener teal. Reshooting was quoted at six weeks. Someone ran the catalog through an AI recolor batch job over a weekend and uploaded results Monday. By Wednesday, legal flagged three assets. Two models wore trademarked team colors the recolor had pushed into competitor territory. One leather bag had magenta fringing along stitch lines where the model hallucinated a hue boundary. Marketing loved the speed. Brand compliance hated the deltaE.

AI recoloring is a legitimate production shortcut. It is not a brand approval pass, a retouching license, or a replacement for shooting product in the correct colorway. Generative tools replace color regions while attempting to preserve luminance texture, shadows, and material cues. They do not apply a CSS filter. They predict new pixel values conditioned on a target palette or prompt. The prediction fails in predictable places: fine mesh, reflective metal, semi-transparent packaging, and any region where segmentation confidence drops.

This guide covers what recolor tools actually do in Adobe and online services, how to anchor recolors to brand OKLCH values before pixels move, how verification with sampling and edge inspection catches failures batch jobs hide, and why changing hex values at rest does not change trademark law or model releases. Speed and compliance are compatible only when the gates stay in order.

What generative recolor changes in the pixel pipeline

Recolor capabilities appear across Adobe Creative Cloud and a fragmented category of web-based batch services. Photoshop documents generative features including selection-assisted recolor workflows that shift apparel and packaging hues while retaining fold shadows. Illustrator Generative Recolor targets vector artwork and flat regions, useful for mockups before photo work. Adobe Firefly sits in the broader generative stack with commercial terms on adobe.com/legal/licenses that vary by subscription tier and change over time. Check live terms before batch commercial deployment.

Capabilities generally include prompt-driven palette suggestions, reference color sampling, and iterative variation. Limitations generally include edge bleed on fine detail such as stitching and hair, metamerism shifts under different white balance, and unpredictable saturation on specular highlights where color is mostly reflection rather than pigment. Online services offer upload-and-download batch flows with varying mask support. None share a standard API contract or brand-governance layer. Treat uploads like any SaaS processing catalog imagery. Read commercial use, retention, and training rights on uploads. Assume files may live on vendor servers. Do not send unreleased products if NDAs prohibit it. Export provenance with original file hash, service name, date, and operator account.

Recoloring modifies pixel values. It does not automatically grant rights to use stock photos outside license terms, model release coverage for altered likeness or uniform colors, trademark clearance when a jersey shifts toward another team’s palette, or indemnification when tool terms disclaim commercial liability. You still need chain of title on the source photograph. Recolor is derivative work. Legal review applies for celebrity talent, sports licensing, and co-branded packaging. Recolor also does not replace product truth-in-advertising obligations. If the physical SKU ships in navy and the recolor shows teal, commerce policy blocks publish regardless of color science.

The distinction between palette exploration and brand commitment shows up inside Adobe workflows when Generative Recolor suggests harmonious companion colors. That is exploration, not semantic token assignment. Companion colors still need contrast checks if they appear behind typography in marketing composites. The photography target sheet and the CSS token file should share one source of truth when brand anchors live in DTCG 2025.10 JSON as oklch components.

OKLCH anchors before anyone moves a pixel

Hex from a PDF brand guide is a starting point, not the control variable. Convert anchors to OKLCH per CSS Color Module Level 4 and Björn Ottosson’s Oklab definition, the same space browsers use in oklch(). Build a one-page recolor target sheet with roles, OKLCH coordinates, sRGB hex fallbacks for tools that only accept hex input, and tolerance thresholds for verification sampling.

--brand-primary: oklch(0.55 0.14 195);
--brand-secondary: oklch(0.72 0.08 195);
--brand-accent: oklch(0.58 0.18 45);

Feed tools hex if required, but verify in OKLCH afterward. Recolor engines often operate in RGB or LAB internally. Hex round-trip can shift hue angle on teals and purples where small numeric changes read as large perceptual differences. Tolerances should be expressed in OKLCH deltas, not in hex digit changes, because hex hides cylindrical geometry.

Segmentation and masking deserve as much attention as anchors. Auto-segmentation is fast and sloppy on fine mesh, lace, chain links, semi-transparent packaging, reflective metal, and logos that must remain frozen including team marks and certification badges. Manual mask refinement beats regenerating entire scenes. Exclude skin unless you have explicit approval to alter skin tone. Exclude text overlays because OCR labels should not be recolored by diffusion models. For apparel, mask the garment layer only and preserve background sweep neutrality so sitewide surface tokens do not fight a pink-gray cast introduced by global recolor.

Process one hero SKU manually before batching forty variants. Start from the sampled hex closest to your OKLCH anchor. Request lower saturation first and increase chroma only after edges pass. Save non-destructive layers or versioned exports. Do not prompt make it pop without numeric targets. Do not accept the first of four variations because the thumbnail looked fine. Do not batch recolor logos, patches, or licensed marks. Do not assume white-balance correction is included. Shoot gray card reference when possible so later verification has a neutral baseline.

Our image color picker samples regions from exported stills and reports converted values. It belongs on the verification side after recolor, not as the sole generative input, because sampling flat regions at one hundred percent zoom against anchor tolerances is how you catch hue drift batch thumbnails hide.

Forty SKUs over a weekend and the rollback that followed

The catalog crisis landed at a mid-market outdoor retailer during a spring rebrand that moved primary from cobalt toward a greener teal positioned as sustainability-forward without abandoning heritage navy equity. Forty core SKU photos on neutral gray sweeps needed to align with the new story for a Memorial Day campaign. Studio reshoot quotes landed between five and six weeks with weather risk for lifestyle exteriors and no studio capacity for pack shots until after launch. The director of ecommerce approved an AI recolor pilot with explicit legal review on any asset containing logos, talent, or co-branded patches.

A contractor ran Saturday and Sunday through a Photoshop-assisted workflow on twelve heroes first, then a web batch service for the remaining pack shots after heroes passed a quick director review that was visual, not numeric. Monday upload put recolored assets into the CMS staging environment. Marketing celebrated the timeline save in a Slack thread with before-and-after grids. Brand compliance had not been in the loop on Sunday night.

Tuesday morning compliance opened a hero jacket frame at two hundred percent zoom and found cyan fringing along zipper tape where segmentation had been loose. The flat chest panel sampled close to target teal, but the zipper edge averaged toward a hue that clashed with the brand sheet. A second hero showed leather bag stitching with magenta micro-fringing classic generative boundary error. Compliance rejected two of twelve heroes immediately and expanded spot-check sampling from twenty percent to one hundred percent for anything worn on body or showing hardware edges.

Wednesday legal review focused on assets the recolor had not broken technically but had broken contractually. Two images showed athletes in licensed team apparel. Recolor had shifted jersey accents toward hues legal identified as uncomfortably close to a rival team’s trademark palette in the same league. The model releases and league marketing agreements did not contemplate hue alteration. Even if the pixels looked better to marketing, the derivative work category changed risk. A third asset showed a co-branded patch that the batch tool had partially recolored despite mask instructions because the web service’s auto-segmentation had included the patch region in the garment class.

The rollback removed twenty-three assets from staging, not forty, because seventeen solid-color pack shots on neutral sweeps passed both OKLCH flat-region sampling and edge inspection after a second pass with tighter masks. The company published on time by pairing recolored solid SKUs with a smaller set of reshot heroes booked in emergency studio time for the two jacket and bag frames and three licensed apparel images replaced with SKU-only angles from an older shoot whose licenses clearly covered color reference without talent faces.

The postmortem did not ban AI recolor. It mandated OKLCH anchor sheets signed by brand before batch runs, one hundred percent review on heroes and anything with logos or skin, legal pre-clearance on licensed marks before segmentation, and provenance logs archived with generative tool terms for the batch date. Marketing could still move fast. Compliance moved earlier in the timeline instead of after upload.

The seventeen assets that survived taught a quieter lesson. Solid-color SKUs on neutral sweeps with manual masks excluding hardware and labels recolored within tight OKLCH tolerances reliably when operators lowered saturation first and increased chroma only after flat regions passed. Iridescent nylon, holographic foil, fine tartan, and angle-dependent sportswear still belonged in reshoot or physical sample photography. The pilot’s success was partial and conditional, which is the honest outcome most catalogs should expect.

Verification, page harmony, and licenses that recolor does not rewrite

Verification starts on recolored masters at one hundred percent zoom. Sample flat fill regions such as chest panels, bag faces, and shoe uppers, not specular dots. Convert samples to OKLCH and compare against anchor tolerances expressed as deltas in L and C with hue checked separately because small hue drift on neutrals matters less than on saturated teals. Shadow regions may read higher L than anchors without failing if the failure mode is neon shadow, not darker shadow. Highlights near white should not force hue where the material is specular.

Edge artifact checks at two hundred to four hundred percent zoom catch halo and fringe where segmentation was loose, texture smear where leather grain melted into plastic, pattern bleed where plaid threads merged across hue cells, and JPEG block edges amplified by recolor compression. Document failures with cropped screenshots in asset tickets. Approve only when both flat OKLCH and edges pass. For large batches, sample one hundred percent of heroes and campaign assets, twenty percent random SKU stills when the first pass is clean, and one hundred percent of anything with logos, text, or skin. Failed spot-checks expand sample size to the whole batch.

Recolored photos must harmonize with CSS tokens on the product page. A teal product on a surface token with competing hue temperature reads as clash even when the SKU sample matches brand OKLCH. Composite recolored cutouts on exported surface hex adjacent to accent CTA buttons before publish. Check that product hue does not mirror danger or success semantics in the parent system. When sourcing new shots, grade candidates against the same OKLCH anchor sheet before they enter the recolor pipeline. Prevention beats another generative pass.

Web delivery should embed sRGB for heroes and thumbnails in most ecommerce grids. Wide-gamut Display P3 exports are rarely needed for catalog grids because sRGB keeps parity with CSS tokens authored in OKLCH with hex fallbacks per CSS Color 4 gamut mapping. Archive masters as PNG or TIFF with documented profiles. Strip generative metadata only if legal approves retention policies. Some teams retain generation logs for audit.

Before batch processing, confirm source photo license whether rights-managed stock, studio contract, or in-house shoot. Confirm generative tool terms for enterprise versus consumer tiers on online services. Confirm whether vendors use uploads to train models and opt out if available. Confirm indemnification limits. Recolor does not create safe harbor for using a competitor’s product photo, a runway image you do not own, or a celebrity editorial licensed for print only. Evaluate services against the legal checklist in AI-Generated Images and Commercial Web Use.

Reshoot or physical sample photography when material is iridescent or angle-dependent, when color accuracy is regulated for categories like food or medical devices, when fine pattern carries the brand such as tartan or logo repeat, when legal requires physical color reference matching lab spectro readings, or when marketplace seller policies reject generative alteration. AI recolor wins on solid-color SKUs, neutral sweeps, and timeboxed rebrands. It does not win on every catalog line.

AI recoloring moves product pixels toward a new palette faster than a reshoot when masks, anchors, and verification gates are respected. Brand safety comes from OKLCH targets signed before batch runs, careful exclusion of skin and licensed marks unless counsel clears them, pixel verification on flat regions, edge QA at magnification, and the same licensing discipline as any derivative image work.

Recolor changes values at rest. It does not change trademark law, model releases, or contrast math on the product page beside the asset. Sample brand OKLCH, recolor with constraints instead of vibes, verify with a color picker on flat regions, inspect edges before upload, clear legal on licensed content, then publish. The gates feel slower than a weekend batch until you have lived the rollback that follows skipping them.

The catalogs that benefit most from generative recolor are the ones that treat it as a conditional tool in a larger asset system, not as a universal filter. Solid neutrals, disciplined masks, signed anchors, and magnification-based edge review are unglamorous compared with one-click batch buttons. They are also what let marketing keep the timeline while brand compliance signs the release without a crisis meeting on Wednesday.