Still Image Retouching Finishing

Still Image Retouching and Finishing

This skill is for finishing existing still images. It assumes there is a captured, scanned, rendered, or composited source image and that the agent's job is to make production decisions about development, retouching, color, metadata, output, and review. It is tool-agnostic: Lightroom, Camera Raw, Capture One, Darktable, RawTherapee, Affinity Photo, Photoshop, GIMP, command-line processors, DAM systems, RIP software, and lab portals are possible instruments, but the skill's decisions are about craft and delivery rather than any one vendor.

Use it for:

• RAW intake, rendered-file intake, selects, versioning, sidecars, and handoff notes.
• Nondestructive development: exposure, white balance, tone mapping, lens corrections, local masks, snapshots, proof copies, layers, smart objects, adjustment stacks, and reversible repair.
• Color-managed finishing: working spaces, embedded profiles, monitor profiling, ICC soft proofing, rendering intents, output profiles, print proofing, and web/social conversions.
• Retouching: dust, sensor spots, blemishes, flyaways, fabric wrinkles, product cleanup, architecture perspective, food freshness, documentary limits, compositing cleanup, and client/legal review.
• Generative fill or AI repair decisions only as post-production interventions on existing stills, including disclosure, provenance, and rights notes.
• Sharpening, noise reduction, resampling, crop variants, batch consistency, metadata, accessibility metadata, export recipes, QA, and archive/delivery packages.

Do not use it for:

• Generating new images from prompts as the primary task.
• Choosing or integrating image-generation APIs.
• Video color grading, motion retouching, VFX shot work, or frame-by-frame video finishing.
• Campaign concepting, art direction, casting, styling, media buying, or platform growth strategy except where those affect still-image finishing constraints.

Grounded Facts

These are documented facts, not taste preferences.

• ICC profiles describe device or color-encoding behavior so color management systems can transform color data between native device color spaces and device-independent spaces. ICC v4 uses a Profile Connection Space based on CIE colorimetry; ICC identifies input, display, and output device profile classes and defines rendering intents for gamut mapping. Source: International Color Consortium, "Introduction to the ICC profile format" and "ICC Specifications," verified 2026-07-11, https://www.color.org/iccprofile.xalter and https://www.color.org/icc_specs2.xalter.
• ICC.1:2022 profile version 4.4.0.0 is the current ICC v4 specification listed by ICC, and iccMAX is specified separately as ICC.2:2023 / ISO 20677 for extended needs such as spectral data or modified PCS behavior. Source: ICC Specifications, verified 2026-07-11, https://www.color.org/icc_specs2.xalter.
• CIE's electronic International Lighting Vocabulary provides public access to terms and definitions from CIE S 017:2020. Use CIE language for colorimetry, illuminants, observers, luminance, reflectance, and viewing-condition claims rather than inventing casual definitions. Source: CIE e-ILV, verified 2026-07-11, https://cie.co.at/e-ilv.
• Lightroom Classic official documentation states that Develop previews use ProPhoto RGB by default, other modules primarily use Adobe RGB, and export/print can choose output profiles such as sRGB or custom printer profiles. It also states that soft proofing previews destination appearance and gamut warnings, and that proof copies can preserve the original edit state. Source: Adobe Lightroom Classic color management and Develop module documentation, verified 2026-07-11, https://helpx.adobe.com/lightroom-classic/help/color-management.html and https://helpx.adobe.com/lightroom-classic/help/develop-module-options.html.
• Camera Raw and Lightroom process versions affect available controls and rendering. Adobe documents newer process versions for improved color mixer banding behavior, high-ISO rendering, dehaze, range mask behavior, tone controls, sharpening, and noise reduction. Source: Adobe Camera Raw and Lightroom process-version documentation, verified 2026-07-11, https://helpx.adobe.com/camera-raw/using/process-versions.html and https://helpx.adobe.com/lightroom-classic/help/develop-module-options.html.
• Exif is an embedded metadata family for digital still-camera files built on JPEG/TIFF/WAV-related specifications, primarily recording technical capture information. CIPA and JEITA jointly manage the Exif standard; the Library of Congress described Exif 3.0 as the most recent version in November 2023 and notes that GPS support was introduced in Exif 2.0 with improvements in 2.2. Source: Library of Congress format description, verified 2026-07-11, https://www.loc.gov/preservation/digital/formats/fdd/fdd000618.shtml.
• IPTC Photo Metadata defines professional photo metadata fields through IPTC Core and IPTC Extension, including rights, creators, people, locations, products, accessibility descriptions, and, in IPTC Photo Metadata 2025.1, AI-related properties. Those 2025.1 AI fields target metadata for AI-generated content; they do not by themselves satisfy every client, publisher, platform, legal, or ethical disclosure requirement for AI-assisted retouching, generative repair, or composite finishing. IPTC states that it uses XMP as one technical format and IIM as another. Source: IPTC Photo Metadata Standard, verified 2026-07-11, https://iptc.org/standards/photo-metadata/iptc-standard/.
• XMP is an extensible metadata framework for embedding metadata in files. Adobe's public XMP documentation describes Part 1 as the data model, serialization, and core properties; Part 1 is also ISO 16684-1:2012. Source: Adobe XMP documentation, verified 2026-07-11, https://www.adobe.com/devnet/xmp.html.
• C2PA develops technical standards for certifying source and history, or provenance, of media content. The C2PA specification site includes Content Credentials, implementation guidance, security considerations, and AI/ML guidance. Source: C2PA Specifications 2.2 index, verified 2026-07-11, https://spec.c2pa.org/specifications/specifications/2.2/index.html.
• Adobe documents Content Credentials in Photoshop as a way to attach attribution details and edit history on export. This is implementation-specific; do not treat it as a universal substitute for IPTC rights metadata, delivery notes, or client disclosure. Source: Adobe Photoshop Content Credentials documentation, verified 2026-07-11, https://helpx.adobe.com/photoshop/using/content-credentials.html.
• Adobe documents sharpening and noise reduction as operations that require 100% or 1:1 inspection for accurate judgment; Lightroom sharpening includes Amount, Radius, Detail, and Masking; Smart Sharpen includes Amount, Radius, Reduce Noise, and blur-type removal choices. Source: Adobe Lightroom retouch documentation and Photoshop Smart Sharpen documentation, verified 2026-07-11, https://helpx.adobe.com/lightroom-classic/help/retouch-photos.html and https://helpx.adobe.com/photoshop/using/adjusting-image-sharpness-blur.html.
• The NPPA Code of Ethics says visual journalists should be accurate and comprehensive, resist staged manipulation, not alter events, and maintain the integrity of image content and context; editing must not mislead viewers or misrepresent subjects. Source: National Press Photographers Association Code of Ethics, verified 2026-07-11, https://nppa.org/resources/code-ethics.

Practical Observations

These are production observations that recur across tools and clients. They should be tested against the actual file, brief, and delivery path.

• RAW development is usually the least destructive place to set white balance, global exposure, highlight recovery, shadow placement, lens corrections, chromatic aberration removal, and broad color rendering. Pixel editing is better reserved for repairs, composites, local shape/texture work, and output-specific interventions.
• Rendered files such as JPEG, PNG, HEIC, TIFF, PSD/PSB, or web downloads have already had a demosaic/rendering/color transform applied. Treat them as smaller edit envelopes than camera RAW unless bit depth, profile, and edit history prove otherwise.
• A visually good file can still be production-bad if it lacks an embedded profile, clips important channel detail, contains hidden retouch artifacts, strips required rights metadata, changes documentary meaning, or fails the intended viewing medium.
• A retouch plan is more reliable when it names both the aesthetic target and the truth boundary: "remove temporary lint and sensor dust, keep garment shape and skin texture" is safer than "make it clean."
• Soft proofing is a prediction, not an oracle. It depends on monitor profile quality, ambient viewing conditions, the destination profile, rendering intent, paper/ink simulation, and the lab or press process actually used.
• Output sharpening is destination-dependent. A file prepared for matte fine-art print, glossy magazine reproduction, retina web display, and social recompression rarely wants the same sharpening or noise texture.
• Social platforms often recompress, resize, strip metadata, or convert color. Deliver a clean sRGB version for common social use unless the platform/client explicitly supports a different profile or HDR path, and keep the richer master separately.

Production Heuristics

Heuristics are defaults to reason from, not rules to obey blindly.

1. Preserve originals and edits. Keep the untouched capture/render, sidecar/catalog state, layered master, flattened delivery exports, and notes on any destructive or generative intervention.
2. Decide the truth class before touching pixels. Editorial documentary, commercial beauty, product e-commerce, architecture, food, fine art, and restoration have different permissible edits.
3. Work broad to narrow. Correct profile interpretation, camera/lens issues, white balance, exposure, tone, and global color before local masks, cleanup, liquify/warp, dodge/burn, skin, product edges, and output sharpening.
4. Use reversible structures when available. Prefer sidecars, virtual copies, snapshots, adjustment layers, layer masks, smart objects, clone/heal on blank layers, named groups, and export presets over flattened edits.
5. Judge detail at the right zoom. Evaluate noise, sharpening, dust, skin texture, moire, halos, resampling, and clone artifacts at 100%; evaluate composition, color balance, and emotional read at fit-to-screen and expected display size.
6. Convert late, embed profiles, and proof before delivery. Keep a high-quality working/master state, then export destination-specific derivatives with explicit profile, dimensions, bit depth, compression, metadata, and sharpening choices.
7. Treat metadata as part of finishing. Technical Exif, IPTC rights/creator/caption, accessibility text, content warnings, model/property releases, AI/generative disclosures, and C2PA/Content Credentials may affect whether a file is usable.
8. Batch for consistency, not blindness. Sync exposure/color/lens settings across a set, then inspect edge cases: mixed light, different ISO, lens swaps, skin-tone variance, clipped product colors, and crops with different output ratios.

Intake and Triage

Start by asking what the image must remain true to and where it will go.

Minimum intake checklist:

• Source type: camera RAW, DNG, scan, JPEG/HEIC, layered PSD/PSB, TIFF, PNG, screenshot, generated image, composite, or unknown.
• Bit depth, pixel dimensions, color profile, embedded metadata, sidecar/catalog state, and available original.
• Usage: editorial, documentary, e-commerce, beauty, fashion, food, architecture, real estate, fine art, archive, social, paid ad, print, packaging, internal review.
• Delivery: file formats, size/ratio variants, color profile, print process/profile, paper, web CMS, social platform, lab/press specs, naming convention, deadline.
• Permissions: copyright owner, photographer, model/property releases, location restrictions, artwork/product trademarks, client retouching rules, generative-AI policy.
• Truth boundary: what may be removed, improved, reshaped, reconstructed, synthesized, disclosed, or left untouched.

Classify the asset before editing:

• RAW original: build a nondestructive development recipe, keep the capture as source of truth, and export rendered deliverables only after proofing.
• Rendered original: verify profile and bit depth; avoid excessive tonal recovery; consider requesting RAW/TIFF/master if the requested change exceeds the file's latitude.
• Layered master: inspect layer names, masks, adjustment order, hidden layers, linked assets, embedded profiles, and existing retouch intent before changing it.
• Composite or AI-assisted still: identify ingredients, licenses, disclosure requirements, provenance data, and visible seams; do not present it as a documentary capture.
• Scan/restoration: preserve a raw scan, correct dust/fading/cast carefully, retain historical features unless explicitly restoring a display copy, and record interpretation.

Development Order

A robust still-finishing pass usually follows this order:

1. Ingest and preserve. Duplicate or catalog the source; checksum or otherwise track critical originals when the work is archival or high value.
2. Interpret correctly. Confirm embedded profile, camera profile, RAW process version, demosaic settings, lens profile, orientation, crop, and metadata time/location sanity.
3. Normalize defects. Fix sensor spots, dead pixels, chromatic aberration, lens distortion, perspective issues, moire, and obvious scanning/capture defects before creative grading.
4. Establish neutrals and exposure. Set white balance, exposure, black/white points, highlight rolloff, shadow floor, and midtone contrast according to the subject and output medium.
5. Shape tone and color. Use curves, color mixer, color balance, selective color, calibration/profile choices, or equivalent controls to support the intended look without breaking memory colors or required product matches.
6. Localize with masks. Use subject, luminance, color, gradient, brush, path, frequency, or object masks as needed; feather and refine at natural transitions.
7. Retouch and repair. Heal/clone/patch/dodge/burn/liquify/warp/composite only within the truth boundary; keep repair layers separate when the tool permits.
8. Proof and variant. Soft proof or simulate target conditions, create print/web/social/crop variants, and apply destination-specific output sharpening and metadata rules.
9. QA and deliver. Inspect at multiple zoom levels, compare before/after, check metadata/profile/export specs, and package master plus derivatives according to the handoff.

Color and ICC Decisions

Do not use color-management terms as decoration. Make explicit decisions.

Working state:

• For RAW development, stay in the tool's high-gamut internal pipeline and avoid premature conversion to sRGB.
• For layered retouching, choose a working RGB space compatible with the source, output, and collaborators. Common choices include ProPhoto RGB for high-gamut photo work, Adobe RGB for print-oriented interchange, and sRGB for web-only or low-risk rendered assets. A larger space is useful only if the pipeline preserves enough bit depth and every collaborator understands the profile.
• Preserve or embed the correct profile when exchanging files. An untagged RGB file is ambiguous.

Display/proofing:

• Calibrate/profile the monitor with a measuring device when color matters; keep ambient light stable and avoid strong colored surroundings.
• Soft proof against the actual destination profile when available: printer/paper/ink profile, CMYK press profile, lab profile, or constrained web/social profile.
• Use gamut warnings as diagnostic signals, not automatic commands. Decide whether hue, saturation, contrast, or perceptual relationship matters most.

Rende

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: calesthio
• Source: calesthio/generative-media-skills
• License: MIT
• Homepage: https://github.com/calesthio/OpenMontage

Install and usage instructions live in the source repository linked above.