Install any skill in seconds. Free to start, no credit card required.
Get Started Free →moskoo/xxg-portrait-rebuild-light
Edit an existing JPG, JPEG, PNG, or WebP portrait to rebuild physically coherent light, exposure, color, capture style, and clean optical skin realism without changing the person.
Passed 19 of 23 test cases with the skill, and 13 of 23 without it.
and 3 more in the full report
6 fixed14 passed either way4 still failing0 broke
Tokens, turns and time describe resource use, not a dollar price.
This benchmark measured an earlier version of the skill.
One of 5 test cases the skill fixed. Same model, same prompt, with and without the skill.
When invoking the native Codex image generation callable image_gen__imagegen for an image-to-image portrait edit, what is the exact argument name used to supply the path to the local source photo?
| Checked | Without skill | With skill |
|---|---|---|
| The tool parameter name for the source image path is referenced_image_paths. | Missed | Passed |
The instructions your model reads when the skill loads.
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:
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.
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.
ALL_TOOLS and prefer the exact discovered name image_gen__imagegen.const result = await tools.image_gen__imagegen({
referenced_image_paths: ["/absolute/path/source.png"],
prompt: "compact English image-edit prompt"
});
generatedImage(result);
test cases passed, vs 13 / 23 without
More skills from the same source and nearby in the registry, with their headline benchmark lift.