Visual-review checkpoint — survey-derived L2 looks + skin primitives

Companion to v1.10's photographer-workflows survey work (RFC-031/032/033 + scope-expansion). This is the darktable-session validation pass for entries that ship without lab-grade automated coverage — the natural-skin-vs-painted-skin / dramatic-vs-cartoonish / restrained-vs-flat boundaries are visual judgments, not assertions. When · Run before closing RFC-033 into ADR or before any production photographer session that depends on these entries.

What needs visual sign-off

Skin primitives (RFC-033 + cheap-#4 scope expansion):

• skin_uniformity — Path B color-band uniformity via colorequal sat_orange shift, masked to mask_skin_region
• skin_smooth_painterly — Frequency-separation approximation via masked clarity reduction

Survey L2 looks (14 new):

• 5 portrait: look_portrait_natural_skin, look_portrait_editorial, look_portrait_skin_warm_lift, look_portrait_background_dim, look_portrait_split_tone_moody
• 9 landscape: look_landscape_grand_vista, look_landscape_intimate_quiet, look_landscape_golden_hour, look_landscape_blue_hour_cool, look_landscape_atmospheric_haze, look_landscape_dramatic_moody, look_landscape_autumn_pop, look_landscape_sky_enhance, look_landscape_water_silk

Required raws

• 3-5 portraits with varied skin tones (caucasian-through-deep), varied lighting (window light, studio strobe, golden-hour, mixed), and varied makeup load (natural through full editorial). At least one with visibly uneven skin (red patches on forehead, yellow undertones on cheeks).
• 5-7 landscapes covering: grand-vista wilderness, intimate small-scene (forest interior, abstract detail), golden-hour scene, blue-hour / twilight, hazy / misty / fog, dramatic stormy weather, autumn foliage, water surface (long-exposure or glassy lake).

Per-entry checklist

For each entry, render (render_preview or darktable-cli with the bundled baseline XMP) and visually verify the intent matches the result. Score one of:

• ✅ Matches — the rendered image reads as the entry's stated intent
• ⚠️ Approximate — generally right but parameters need tuning (note suggested adjustment)
• ❌ Wrong — fails the intent (note what's wrong; consider re-design or fall-back path)

Skin primitives

skin_uniformity at strengths 0.0 / -0.3 / -0.6

Test image	Strength	Expected	Pass?
Portrait with red forehead patches	-0.3	Patches noticeably reduced; natural skin still
Portrait with red forehead patches	-0.6	Strong uniformity; check it doesn't read as "painted"
Portrait with deep skin tones	-0.3	Uniformity works without lifting tone reading as flat / muddy
Portrait with full lipstick/makeup	-0.5	Lip color preserved (mask_skin_region's color_h fallback should exclude lips); if lipstick desaturates badly, that's a mask-leakage finding
Studio-strobe portrait	-0.5	Skin reads natural under hard light (the failure mode here is "plastic skin")

Fail criteria: if skin_uniformity produces "plastic skin" at strength -0.5 on representative portraits, the visual quality bar fails. Fall back to RFC-033 Path A (full Path C decoder of colorequal — substantially heavier work) or document the strength ceiling.

skin_smooth_painterly at strengths -0.3 / -0.5 / -0.7

Test image	Strength	Expected	Pass?
Close-up portrait with visible pores	-0.3	Subtle smoothing; pores still visible
Close-up portrait with visible pores	-0.5	Noticeable smoothing without crossing into "plastic"
Close-up portrait with visible pores	-0.7	Strong smoothing; check this is the upper acceptable bound
Editorial / fashion portrait	-0.5	Composes with skin_uniformity for a complete skin pass

Fail criteria: if -0.5 already produces over-smoothed / waxy skin, narrow the recommended range to -0.3..-0.5 and update the description's "typical -0.3 to -0.7" guidance.

Portrait L2 looks

look_portrait_natural_skin

• Skin tones unchanged in the warm direction (subtle, not orange)
• Contrast feels gentle — not flat, not punchy
• Saturation pulls back without the image looking dull
• Composes naturally with skin_uniformity / skin_smooth_painterly

look_portrait_editorial

• Stronger contrast than natural_skin; visibly punchier
• Split-tone reads (cool shadows + warm highlights) without skin tones turning unnatural
• Saturation pull-back doesn't make image look washed out

look_portrait_skin_warm_lift (masked)

• mask_skin_region binding scopes the lift to skin only
• Background / clothing exposure unchanged
• Skin reads warmer + brighter without the boundary being visible (mask edge feathering OK)
• If the mask leaks onto orange backgrounds / lipstick, document and consider override-mask recipe

look_portrait_background_dim

• Caller-supplied subject mask shape works (no pre-baked mask)
• Background dims by ~0.4 EV; saturation reduction visible but not aggressive
• Subject preserved by the inverse mask

look_portrait_split_tone_moody

• Stronger split than editorial — cool blue shadows + warm orange highlights
• Skin tones survive the strong split (the failure mode: skin turns into the highlight color)
• If skin breaks at the default values, document strength override recipe

Landscape L2 looks

look_landscape_grand_vista

• Sky and foreground both readable; no extreme tone compression
• Mild clarity bite gives definition without crunch
• Warm shadows shift visible without being over-warm

look_landscape_intimate_quiet

• Reads as less processed than the baseline — the explicit Marino-restraint signal
• Soft clarity (negative bilat) gives painterly feel
• If image looks just "blurry," not "painterly," recipe needs sharper tuning

look_landscape_golden_hour

• Warm shift visible but credible (not orange-sunset cartoon)
• Highlights amber; shadows warm; image reads as "this captured the warmth I felt"
• Already-warm scenes don't blow out into pure orange

look_landscape_blue_hour_cool

• Cool shift visible but credible
• Mood reads as twilight / pre-dawn
• If already-cool scenes go too cyan, consider tempering temperature shift

look_landscape_atmospheric_haze

• Hazeremoval lifts visibility WITHOUT killing the moody fog
• Strength 0.5 in the dtstyle is right (not 1.0 or 1.5 — those defeat the intent)
• Composes with golden_hour for warm-hazy scenes

look_landscape_dramatic_moody

• Strong contrast (1.7) reads as dramatic, not crunchy
• Cool shadows + warm highlights split adds depth without skin-tone-style cartoon
• Stormy / weather images: mood emerges; calm scenes: image reads over-processed (which is the right failure mode — this look isn't for calm scenes)

look_landscape_autumn_pop

• Orange / red foliage visibly lifted without the "Lightroom autumn preset" cartoon look
• Sky stays natural (the sat_blue compensation working)
• Non-foliage scenes: applies regardless but reads weird (which is the right failure mode — this look isn't for non-foliage)

look_landscape_sky_enhance (masked)

• mask_sky binding scopes the enhancement to sky only
• Foreground / land unchanged
• On complex skies (sunsets, partial clouds, trees in sky), parametric fallback may leak — escalate via Pattern 7 of llm-vision-for-masks.md and re-test
• Composes on top of grand_vista / dramatic_moody for sky-emphasis

look_landscape_water_silk (masked)

• mask_water_blue_cyan binding scopes the smoothing to water only
• Smoothing actually reads as silky water (not "blur applied to a region")
• Rocks / shore unchanged
• Cyan-tinted backgrounds may leak — note as a finding

How to run a render

# For each test image, ensure a workspace is bootstrapped:
chemigram ingest /path/to/test/raw.NEF --image-id test_portrait_caucasian_window

# Apply the entry under test:
chemigram apply-primitive test_portrait_caucasian_window \
  --entry skin_uniformity \
  --param sat_orange=-0.5 \
  -p expressive-baseline

# Render the preview:
chemigram render-preview test_portrait_caucasian_window
# → outputs JPEG path; open in your viewer of choice

# Visually compare against:
# - the unprocessed baseline render
# - your reference editor (Lightroom / Capture One) doing the equivalent move

For the masked entries, also test: - The pre-baked named mask path (default — apply with no override) - An LLM-vision-constructed mask via Pattern 7 of llm-vision-for-masks.md

What to record

Per entry, capture:

1. Pass/fail/approximate rating
2. Test images used (anonymized identifiers)
3. Specific failure modes observed (e.g., "lipstick desaturates at strength -0.7 on portrait_caucasian_makeup")
4. Recommended adjustments (parameter range narrowing, mask refinement, default value changes)

Outputs

After the review session:

• Failures → file follow-up issues, mark the entry's manifest description with caveats (or remove and re-author)
• Strength-range tuning → update the manifest's range / default fields, regenerate via scripts/generate-l2-looks-survey.py
• Mask-leakage findings → either tighten the maskdef's parametric fallback (RFC-032), document the override-mask recipe in vocabulary-patterns.md, or escalate to LLM-vision construction (llm-vision-for-masks.md Pattern 7)
• Successes → close the visual-review checkpoint and ship the closing ADR for RFC-033 (currently blocked on this checkpoint)

See also

• docs/photographer-workflows-survey.md — the source survey that informed all 14 looks + 2 skin primitives
• docs/rfc/RFC-033-skin-tone-uniformity.md — recommends this checkpoint before ADR closure
• docs/guides/llm-vision-for-masks.md — Pattern 7 for content-aware mask escalation
• docs/guides/vocabulary-patterns.md — composition recipes referencing these entries
• docs/guides/visual-proofs.md — automated visual-proof sweep (different scope: lab-grade vs visual-review)