Xxg Portrait Rebuild Light

XXG Portrait Rebuild Light V2.3

Objective

Treat the input as the same photograph, never as a reference for a replacement portrait. Improve illumination, exposure, color rendering, capture character, skin response, and generation quality while retaining identity, facial geometry and natural asymmetry, expression, pose, camera view, framing, and subject scale. Preserve source optics unless the user explicitly requests an optical restyle.

Build realism from three separable signals:

1. clean low-frequency skin tone and broad transitions;
2. source-driven diffuse/specular response with highlights confined to plausible light-facing areas;
3. fine, region-specific microdetail limited by face scale, focus, and illumination.

Do not manufacture realism with dirt, darkness, coarse pores, uniform grain, random color patches, or exaggerated facial lines. The result must first read as a clean photograph at normal size.

Choose the Edit Scope

• texture-only: when the user asks only to remove plastic/AI skin or recover detail, force L0 + E0 + T0 + G0 + D0 + A0 + Q0. Preserve the source lighting, highlight placement, exposure, white balance, focal plane, depth of field, and background.
• tone-and-exposure: when the user asks for brightness, highlight/shadow recovery, white balance, palette, or grading without new lighting, preserve the source light direction and edit only E/T/G.
• relight-and-skin: when the user requests a lighting change or selects L/T/A, authorize the requested light response while preserving source focal plane and depth of field unless explicitly changed.
• capture-style: when the user requests a camera, film, phone, CCD, or era look, edit G/D and only the exposure behavior logically required by that capture; preserve viewpoint, crop, focal plane, and depth of field.
• optical-restyle: only when the user explicitly requests a different focal-length perspective, camera angle, aperture behavior, or depth of field. State that axis once and permit only the minimum framing/background reconstruction it requires.
• quality-repair: when the user asks to remove dirty synthetic patterns, banding, posterization, seams, halos, or other generation residue, edit only Q and preserve light, color, structure, optics, and natural texture.
• base-color-repair: when an edited result has an unintended cast, use the original root base to restore only unchanged or unauthorized color relationships; preserve deliberate recolors and requested E/T/G/A changes.
• multi-round-repair: when repeated edits accumulate tint, noise, sharpening, compression, banding, or seams, use the latest accepted edit as the target and the first root base as the stable identity/geometry/color reference.

Never let texture repair become relighting, tone correction become a new light source, device emulation become a new pose/viewpoint, or quality cleanup become global smoothing or an unauthorized regrade.

Use the Host Image Editor

1. Inspect the source and read the host's native image-generation or image-editing skill.
2. Discover the actual callable in the tool registry. In Codex, inspect ALL_TOOLS and prefer the exact discovered name image_gen__imagegen.
3. For a local source in Codex, use the discovered tool and its real arguments:

const result = await tools.image_gen__imagegen({
  referenced_image_paths: ["/absolute/path/source.png"],
  prompt: "compact English image-edit prompt"
});
generatedImage(result);

Never guess tools.image_gen or input_image. Correct wrong members, arguments, or TypeError from the registered signature and retry. In Claude, OpenClaw, or another host, use the equivalent native image-edit action explicitly exposed by that host.

For base-anchored or multi-round repair, pass exactly [CURRENT_EDIT, ROOT_BASE] and state: Image 1 is the edit target; Image 2 is the original root reference for declared locked axes only. Do not attach every intermediate round.

Route the Outcome

Reading a skill, inspecting an image, creating a task, or announcing generation is not an image-tool invocation.

Never Produce the Final Image Locally

Do not use Pillow, NumPy, OpenCV, ImageMagick, FFmpeg, sips, or custom raster scripts to relight, grade, retouch, sharpen, add texture, resize, crop, extend, composite, repair, or produce the delivered image. Use them only for read-only aspect-ratio, mask, and result audits. See requirements.txt.

Compile the Image Prompt

Read the V2.3 prompt compiler, the lighting and skin recipes, the tone/exposure/style/device recipes, and the quality/iteration recipes. Decide internally as:

Scope → Key L → Exposure E → Fill/Shadow → Skin scale S → Skin finish P → Light color T → Look G → Capture D → Background/Atmosphere A → Quality Q

Select exactly:

one L + one E + one S + one P + one T + one G + one D + zero or one A + one Q

Use one key-light system. Atmosphere and skin reflections must inherit its direction, size, falloff, and color. E controls exposure, G controls palette/curve, and D controls capture response; none may invent another light. A6 is the sole override: force E6 silhouette exposure, remove all subject fill/catchlights/internal illumination, use L/T only for the rear source and background, and suppress S/P. Q may clean the background but must never restore detail inside the silhouette.

Send four core lines, adding RENDER for a requested grade/device response and QUALITY for an observed repair target:

EDIT: scope, identity/structure lock, and source-optics lock.
LIGHT: one L, one E, physical shadow/background response, T, and optional A.
RENDER: one G plus one D, only when either differs from source.
SKIN: one scale-aware S behavior plus one source-consistent P finish.
QUALITY: one Q repair, only when Q differs from source.
AVOID: only the two or three failures most likely for this source.

• Use four lines when G0 + D0 + Q0; add RENDER only for a requested grade/device and QUALITY only for an observed repair target. Target 55–110 English words, with an absolute ceiling of 135 for combined edits.
• State identity once. Treat the source itself as the identity card; do not invent a new age, personality, beauty description, lens, or aperture.
• Use positive, observable photographic behavior before negative constraints. Omit recipe codes, audits, confidence, backend notes, and reasoning.
• Compile bare words such as cinematic, editorial, HDR, film, DSLR, medium format, or smartphone into visible exposure, palette, tonal, microcontrast, sharpening, and dynamic-range behavior.
• Default to E0 + G0 + D0 + Q0. Never mix multiple device profiles, color looks, or quality recipes. Never claim exact manufacturer color science or pixel identity from prompt wording.
• Keep default prompts free of realism-by-dirt terms: freckles, blemishes, blackheads, rough skin, color irregularity, film grain, gritty texture, under-eye lines, and high contrast. Preserve source-specific marks without naming or amplifying them.
• Use deep, near-black, or hard contrast only when the user explicitly selects backlight, hard light, low-key, neon, or silhouette behavior.
• On retry, replace the failed line instead of appending more instructions.

Every handoff must use this structure with no placeholders.

Preserve the Identity Signature

Keep six source-defined groups stable: face outline/proportions; feature spacing, shape, and size; hairline, parting, and hair mass; source-identifying skin anchors; makeup/accessories; and apparent age/expression. Do not describe these groups in detail to the image model unless a real failure requires a shorter identity retry. Detailed identity checks belong in validation, not the generation prompt.

Under A6, internal features are intentionally hidden. Judge identity from hair/head/body outline, head-to-body ratio, pose, position, and framing.

Apply Optical Skin Realism

• Keep overall skin color clean and continuous across face, ear, neck, and visible upper chest. Local transitions should be gentle and source-consistent, never patch-like.
• Use a diffuse base with small bounded specular highlights only on planes facing the selected key. Avoid a whole-face gloss layer and avoid removing all highlights.
• Vary surface detail by region: cheek pores softer, nose pores slightly clearer, lip texture separate, eye-area structure undisturbed. Do not tile one pore pattern across the face.
• Match detail to the source focus plane, depth of field, face size, and illumination. Never sharpen the whole face, every hair, clothing, and background equally.
• Preserve source-existing marks as identity anchors, but do not list or generate new imperfections by default.
• Preserve source skin tone; do not use fair, whiter, or beauty-grade language unless the user explicitly requests a complexion change.