Skip to content

Chemigram capability survey — what a user can actually do today

Honest, unvarnished snapshot of every photographic move and workflow operation chemigram supports as of v1.10.0 (released 2026-05-10). Two purposes: (1) baseline for planning what to add next, (2) read-from-cold reference for new contributors / agents asking "is X in scope?"

This document organizes capabilities by what a photographer wants to do, not by code structure. Each section lists what's there, what's missing, and where the gap is. Honest about both.

Inventory (post-v1.10.0): Vocabulary loaded: starter (2 entries) + expressive-baseline (112 entries) = 114 vocabulary primitives. 40 parameterized (RFC-021 / RFC-022 / RFC-035; ADR-077..081 + ADR-088). 19 darktable modules have at least one shipped entry. MCP tools exposed: 27 (added apply_per_region_mixed shape via ADR-089, propagate_state per ADR-090, wb_from_gray_card). CLI verbs mirror the MCP surface verb-for-verb plus diagnostic chemigram status (per RFC-020).

1. Tone and exposure

What's there

Global exposure (RFC-021 parameterized — v1.6.0) - exposure --value V — continuous EV in [-3.0, +3.0]. Replaces the previous discrete expo_+0.5 / expo_-0.5 / expo_+0.3 / expo_-0.3 / shadows_global_+/- ladder. Any value the photographer wants.

Tone curve (sigmoid) - sigmoid_contrast (parameterized; range [0.5, 5.0]; 1.0 = mild s-curve, 1.5 = darktable default, 2.5 = aggressive). Replaces the v1.5.x contrast_low / contrast_high ladder. - blacks_lifted, blacks_crushed — pull or crush deep shadows via sigmoid target - whites_open — extend white target

9-band tone equalizer (RFC-022 Tier 2) - toneequalizer (parameterized, 9 axes; each in [-2.0, +2.0] EV, default 0.0). Bands: noise, ultra_deep_blacks, deep_blacks, blacks, shadows, midtones, highlights, whites, speculars. Pass any subset; unspecified bands stay at 0. Closes the "tonal-zone curves" gap previously listed under "What's missing".

Local tone (drawn-mask exposure) - gradient_top_dampen_highlights — top-half EV reduction - gradient_bottom_lift_shadows — bottom-half EV lift - radial_subject_lift — center ellipse +0.6 EV - rectangle_subject_band_dim — middle horizontal band -0.3 EV - Or any EV through an ad-hoc mask: apply-primitive --entry exposure --value 0.7 --mask-spec '<json>' composes parameter values with drawn-form geometry on a per-photograph basis.

What's missing (fundamentals)

  • Manual tone curve (tonecurve module): no per-control-point manual curve. Closest substitute is toneequalizer's 9-band per-zone control, which covers most photographic intents without needing per-curve-point control.
  • Sigmoid blacks_lifted / blacks_crushed / whites_open parameterization — these still ship as discrete entries. Each represents a kind of tonal move (different fields), not a magnitude on the same axis, so they may stay discrete by design (per ADR-081 Tier 0 framing).

2. Color and white balance

What's there

White balance (temperature module) - temperature (parameterized; multi-axis: red_coeff, blue_coeff in [0.5, 4.0] each; defaults 1.0). The first multi-parameter parameterized entry. Replaces the v1.5.x wb_cool_subtle. wb_warm_subtle remains in the starter pack as a discrete teaching artifact. - ⚠️ does not honor masks (darktable pipeline-position issue, see mask-applicable-controls.md)

Saturation / chroma / vibrance (colorbalancergb — all parameterized) - saturation_global (range [-1.0, +1.0]; -1.0 → monochrome, +0.5 → strong boost). RFC-021; replaces v1.5.x sat_* ladder. - vibrance (range [-1.0, +1.0]). RFC-022 Tier 2; replaces v1.5.x vibrance_+0.3. Vibrance protects already-saturated pixels — gentler chroma push than saturation_global. - chroma_global (range [-1.0, +1.0]). RFC-022 Tier 2; less saturated-pixel protection than vibrance, more aggressive than saturation_global at equal magnitudes. - hue_angle (range [-180.0, +180.0] degrees). RFC-022 Tier 2; rotates every pixel's hue around the color wheel.

Brilliance — per-zone luminance shaping (colorbalancergb — all parameterized; #86) - brilliance_global, brilliance_highlights, brilliance_midtones, brilliance_shadows (each range [-1.0, +1.0]). Per-zone luminance shaping — global moves all zones; per-zone variants target specific tonal ranges.

Per-zone chroma - chroma_boost_shadows, chroma_boost_midtones, chroma_boost_highlights — boost color intensity in a tonal zone (not selective on hue)

4-way grade (color tinting per zone) - grade_shadows_warm, grade_shadows_cool, grade_highlights_warm, grade_highlights_cool — push tint toward orange or blue in the named zone - No mid-tone grade entries shipped

What's missing (fundamentals)

  • WB tint axis (magenta ↔ green): the temperature module's mv4 struct stores RGB coefficients, not temp/tint. The 2-axis parameterized temperature exposes red_coeff + blue_coeff — that captures warming / cooling but not the green-magenta axis a photographer thinks in. A separate tint parameter with proper coefficient → tint conversion would close this.
  • Selective hue rotation (HSL-style "shift only greens toward teal"): not present. The Tier 2 hue_angle rotates all pixels uniformly, not a specific hue band. colorzones would unlock this — Tier 3 per ADR-081 (24+ params; surface doesn't reduce cleanly to scalar).
  • Mid-tone grade: only shadows + highlights have warm/cool entries; midtones don't. Pure authoring work; no architectural blocker.
  • Selective color (HSL-style "only affect blues"): not present at all. Same colorzones Tier-3 framing as above.

3. Sharpening and detail

What's there

Local contrast (bilateral filter) - bilat_clarity_strength (parameterized; range [-1.0, 4.0]; default 0.0 = no clarity, 1.5 = pronounced). Replaces the v1.5.x discrete clarity_strong (the strength axis only). clarity_painterly remains discrete — different kind of clarity, not strength. - clarity_painterly — softer painterly local contrast

Edge-aware sharpening (RFC-022 Tier 2) - sharpen (parameterized; range [0.0, 2.0]; default 0.0 = no sharpen, 0.5 = subtle, 1.0 = strong, 2.0 = aggressive). Unsharp-mask sharpening; radius preserved at darktable default 2.0 px.

What's missing (fundamentals)

  • Modern darktable sharpening family: diffuse-or-sharpen (reaction-diffusion, darktable's flagship modern sharpener) and equalizer (wavelet-based per-frequency contrast control) are unrepresented. The shipped sharpen covers the unsharp-mask common case but lacks the more advanced shaping these modules provide.
  • Noise reduction: no entries for denoiseprofile, nlmeans, bilat-as-denoiser. Users with high-ISO files have no go-to. Tier 3 per ADR-081 (per-camera profiles violate the cost-shape guidance).
  • Luminance vs chrominance noise control: derivative of the above.
  • Hot-pixel removal, dust spotting: not in scope (no entries; would need separate vocabulary subtype).

4. Texture and grain

What's there

  • grain_strength (parameterized; range [0.0, 100.0]; 8 = grain_fine-equivalent, 25 = medium, 50 = heavy). Replaces the v1.5.x grain_fine / grain_medium / grain_heavy ladder.

What's missing

  • Negative grain / smoothing as an opposite move.
  • Grain coloring (sepia tint within grain) — possible via the module, no entry.

5. Optical / geometric corrections

What's there

Vignette (post-process, decorative) - vignette --value V — parameterized radial corner darkening (RFC-021); brightness in [-1.0, +1.0] (negative darkens corners; positive lifts). Replaces the v1.5.x discrete vignette_subtle / medium / heavy ladder.

Crop (RFC-022 Tier 2) - crop --param cx=V cy=V cw=V ch=V — parameterized crop rectangle in normalized coordinates [0.0, 1.0]; default cx=cy=0.0, cw=ch=1.0 (no crop). First workflow-primitive parameterized entry; aspect-ratio preserved at -1/-1 (free).

What's missing entirely

  • Lens correction (lens module): no entries despite darktable supporting it for thousands of lens profiles. Tier 3 per ADR-081 (lensfun-coupled per-lens-model database lookup; the parameterization shape is "select profile" not "tune scalar value"). A possible non-Path-C path: an L1 binding entry that auto-applies the lens profile based on EXIF (sidesteps Tier 3 framing entirely).
  • Rotation / perspective correction (flip, ashift): not in vocabulary. Pure authoring work — both modules are reasonable Tier 2 candidates (small struct; scalar parameters).
  • Distortion correction (barrel / pincushion): not in vocabulary.
  • Chromatic aberration removal (cacorrect): not in vocabulary.
  • Retouch / spot healing (retouch, spots): not in vocabulary at all. This is the major portrait gap. The shape is unusual — frequency-separation retouching has stroke-based input that doesn't reduce to a per-image vocabulary entry. Probably needs its own RFC for what shape this would take in chemigram.

6. Highlights and shadows recovery

What's there

  • highlights_clip_threshold (parameterized; range [0.0, 2.0]; default 1.0 = darktable default, 0.95 = subtle, 0.85 = strong). Replaces the v1.5.x highlights_recovery_subtle / highlights_recovery_strong ladder.

What's missing

  • Shadow recovery as an explicit single-named primitive: the parameterized toneequalizer --param shadows=+1.0 covers this and more, but a one-liner shorthand (e.g., L2 recover_shadows look) doesn't yet exist.
  • HDR-style highlight + shadow combined recovery as a single named move: again, achievable via toneequalizer with shadows up + highlights down, but no L2 composite shipped for this common use case.

7. Local adjustments (drawn masks)

What's there

  • 4 mask-bound vocabulary entries (gradient × 2, ellipse, rectangle) — all on the exposure module
  • Ad-hoc masking via --mask-spec: any vocabulary primitive can be applied through a drawn mask region at apply time (CLI flag + MCP mask_spec arg). Engine-tested for every primitive.
  • gradient, ellipse, rectangle drawn-form geometries supported in darktable's wire format

What's not supported

  • Brush-drawn masks (brush form): no encoder; per ADR-076 deferred.
  • Path / polygon masks with curves: only sharp-cornered rectangles are encoded (rectangle path with degenerate Bézier control points).
  • Mask combination (multiple shapes per binding via boolean ops): one form per binding currently.
  • Parametric masks (mask by luma / chroma / hue range, no drawn shape): the darktable feature exists; chemigram doesn't expose it as vocabulary.
  • Content-aware / AI masks ("the manta's belly"): Phase 4 work; no provider shipped.

8. Looks and multi-module presets

What's there

  • look_neutral (starter) — exposure + wb_warm_subtle baseline; the original teaching artifact
  • look_portrait — gentle skin-protective composition (exposure +0.2 EV + sigmoid_contrast 1.2 + colorbalancergb saturation_global -0.1 + vibrance +0.2)
  • look_landscape — vibrant dramatic landscape (sigmoid_contrast 2.0 + saturation_global +0.3 + vibrance +0.2 + bilat_clarity_strength 1.0)
  • look_vintage_film — nostalgia / faded film (sigmoid_contrast 1.2 + saturation_global -0.2 + grain_strength 25 + temperature warm shift)
  • B&W trio also serves as look-shaped composition: bw_convert (Rec. 709 neutral), bw_sky_drama (red-emphasis), bw_foliage (green-emphasis)

What's missing

  • Cinematic genre looks: no look_cinematic_teal_orange, look_film_kodachrome, look_film_portra, look_high_key_portrait, look_low_key_portrait, look_moody_dramatic, etc. Each is a multi-plugin composite of existing primitives — pure authoring work, no architectural blocker.
  • Decade/era looks (look_70s_film, look_90s_grain, look_2000s_digital, etc.): same shape — composition of existing primitives.
  • Style transfer: looks are static composites; the project doesn't currently support "match this reference image" workflows.

The four looks shipped today bring the Looks layer past "essentially nothing" — but it remains the thinnest L2 surface relative to L3. Per the project's "starter is small, Phase 2 grows from session evidence" framing, more looks ship from real session pull rather than upfront authoring.

9. Workflow operations (not photographic; pure chemigram)

Versioning (snapshot / branch / checkout / tag / log / diff / reset)

  • chemigram snapshot <image_id> — capture current XMP as content-addressed object
  • chemigram branch <image_id> --name <branch> — create branch at HEAD
  • chemigram checkout <image_id> <ref> — move HEAD to a ref / hash / branch
  • chemigram tag <image_id> --name <name> — immutable tag (no overwrite)
  • chemigram log <image_id> — operation history (newest first)
  • chemigram diff <image_id> <a> <b> — added / removed / changed primitives between snapshots
  • chemigram reset <image_id> — rewind to baseline ref (ADR-062)
  • chemigram remove-module <image_id> --operation <name> — strip all history entries for a darktable module

Render and export

  • chemigram render-preview <image_id> --size N — JPEG preview
  • chemigram compare <image_id> <a> <b> — side-by-side stitched render of two snapshots
  • chemigram export-final <image_id> --format jpeg --size N — high-quality export

Vocabulary discovery

  • chemigram vocab list — all entries (filterable by --tag, --layer)
  • chemigram vocab show <name> — full manifest fields for one entry
  • chemigram log-vocabulary-gap <image_id> — record a missing-primitive observation

Context (tastes / briefs / notes / sessions)

  • chemigram read-context <image_id> — agent's first-turn context (tastes + brief + notes + recent log + recent gaps)
  • chemigram apply-taste-update --content <text> — direct CLI-only taste write
  • chemigram apply-notes-update <image_id> --content <text> — direct CLI-only per-image note write
  • (MCP path: propose_taste_update → confirm_taste_update, two-step)

Bindings (L1/L2 templates)

  • chemigram bind-layers <image_id> — apply camera-specific (L1) and look-baseline (L2) templates onto the current XMP

Workspace lifecycle

  • chemigram ingest <raw_path> — bootstrap a per-image workspace
  • chemigram status — runtime diagnostics

MCP-only (agent-driven)

The CLI mirrors most MCP tools verb-for-verb. Agent-only tools (no direct CLI): - propose_taste_update, confirm_taste_update — the two-step propose/confirm flow - propose_notes_update, confirm_notes_update — same shape for per-image notes

10. Honest assessment of the gap

Where chemigram is strong today

  • Workflow + versioning: snapshot/branch/diff/tag/reset is comprehensive and Git-shaped. Cheap, reversible, content-addressed. This part feels finished.
  • Engine architecture: synthesizer, mask-binding, MCP/CLI parity, taste/brief/notes context, session transcripts — all the plumbing is solid.
  • Parameterized continuous control (RFC-021/RFC-022/RFC-035 + ADR-077..083 + ADR-088): 40 parameterized entries across 18 modules as of v1.10.0 (every magnitude-ladder Phase 4 module + the original 4 Tier 2 expansions + 4 brilliance axes + 9 colorbalancergb Color Grading axes + dehaze + texture + 3 HSL multi-axis entries via colorequal + filmic + denoise + lens + transform + WB Kelvin UX wrapper + bw_convert v2's 8 Adams-school axes + parametric L2 strength via Path B). Single-axis, multi-axis (⅔/4 param), 8-axis HSL channels, 9-axis toneequalizer, 10-axis lens, and 24-axis HSL panel cases all proven on the same architecture. Discoverability via MCP list_vocabulary + CLI vocab show (#89). Total vocabulary corpus: 114 entries (40 parameterized + 74 discrete: L1/L2 looks + L3 kinds + mask-bound) plus 9 named maskdefs (RFC-032 — separate kind, not counted in the 114). Did-you-mean suggestions on unknown entry names (#107). Note: 19 total darktable modules have at least one shipped entry (the 18 parameterized + channelmixerrgb which ships Path A discrete entries bw_sky_drama/bw_foliage).
  • Read-side analytics (v1.9.0): chemigram gap-log sub-app (#106) for vocabulary_gaps.jsonl — list / rank / show / clear, supports the Phase 2 use-driven feedback loop. chemigram session-log sister sub-app (#109) for session transcripts — list / show / find / replay. Both with full guides under docs/guides/.
  • Tone control: exposure (parameterized EV ±3) + sigmoid_contrast (parameterized) + toneequalizer (9-band parameterized) + highlights_clip_threshold (parameterized) + blacks_lifted/crushed/whites_open (discrete kinds) + 4 mask-bound exposure entries. The whole tonal surface is photographically operational.
  • Color grading via colorbalancergb: 8 parameterized axes (saturation_global, vibrance, chroma_global, hue_angle, brilliance × 4) plus 7 discrete entries for per-zone chroma + warm/cool grade.
  • B&W conversion: bw_convert / bw_sky_drama / bw_foliage via channelmixerrgb mv3 — closes #63 and the "channel mixer / B&W" gap from §2.
  • Drawn-mask masking: works on every primitive (engine-tested), with three drawn-form geometries (gradient / ellipse / rectangle).
  • CI gates: 5-layer parameterized coverage (ADR-080), manifest-modversion consistency (#85), runtime modversion drift detection (ADR-082) — three separate safety nets against the failure modes parameterization could introduce.

Where chemigram is thin today (post-v1.9.0 in-progress)

The "thin" list has narrowed dramatically since v1.7.0. v1.8.0 closed the bulk of the Tier 3 module watchlist; the in-progress v1.9.0 cycle has shipped CLI analytics (gap-log + session-log + cache + vocab validate), real-raw fixture path, more vocabulary entries (cinematic L2 looks + L3 discrete kinds + 5 compositional-mask L2 looks), and closed the full mask + retouch architecture trilogy: spatial masks (RFC-029 / ADR-084), parametric range filters (RFC-024 / ADR-085), LLM-vision content-derived masks (RFC-026 / ADR-086), and spot heal/clone (RFC-025 / ADR-087). The deployed-sibling-provider precision tier (RFC-030) is drafted and deferred until LLM-vision precision becomes the bottleneck.

Closed in v1.8.0 (decoder shipped or partial-empirical): - ✅ Selective color HSL — closed via RFC-023 / ADR-083 → colorequal mv4 (3 multi-axis entries; 24 axes total). - ✅ Noise reduction — closed via #96 → denoiseprofile mv12 (NLMEANS mode; wavelet-curve verification under #100). - ✅ Lens correction — closed via #95 → lens mv10 with manual-override strength axes (EXIF auto-binding under #100). - ✅ Diffuse-or-sharpen — closed via #92 → diffuse mv2 (texture entry). - ✅ Dehaze — closed via #90 A.2 → hazeremoval mv3. - ✅ Filmic v6 — closed via #97 → filmicrgb mv6 alongside the existing sigmoid. - ✅ Rotation / perspective — closed via #101 → ashift mv5 (transform entry). - ✅ WB Kelvin UX — closed via #102 → wb_kelvin_delta (UX wrapper on temperature). - ✅ Color Grading completion — closed via #91 + #90 A.4.

Closed / shipped in v1.10.0 (photographer-survey-driven vocabulary expansion + workflow primitives): - ✅ 6-genre photographer-workflow survey (docs/photographer-workflows-survey.md) — 36 photographers across portrait / landscape / wedding / B&W / nature-wildlife / food-product genres; full intent-vs-tool mapping; gap matrix at v1.0. - ✅ 29 new L2 looks drawn from the survey: 5 B&W (look_bw_classic_neutral / _high_contrast_chiaroscuro / _landscape_dramatic / _silver_efex_zone_balanced / _split_tone_warm_shadows) + 9 landscape + 5 portrait + 5 wildlife + 4 food + 1 product. All composites of the parameterized L3 primitives. - ✅ bw_convert v2 — colorequal-based 8-axis Adams-school B&W primitive (8 sat=-1.0 + 8 bright_X parameters per hue band; emulates Photoshop Channel Mixer (Monochrome) + Silver Efex color filters). Replaces the v1.4.0 channelmixerrgb-mv3 bw_convert. - ✅ Parametric L2 strength — RFC-035 / ADR-088 (Path B per-parameter interpolation). chemigram apply-primitive --entry <look> --strength 0.5 dials any L2 look from 0 (identity) to 1.0 (authored). Identity values from the parameterize registry; no per-look manifest declarations. - ✅ Mixed-op apply_per_region — RFC-036 / ADR-089. Each region carries its own ops array; per-(op, region) multi_priority allocation. Eye-detail / face-sculpt-with-clarity / sky-and-foreground twin moves now ship as one snapshot. - ✅ propagate_state MCP verb — RFC-037 / ADR-090. LR-Sync analog: anchor on one image, propagate edit state to N targets with framing-bound auto-exclusion (ashift / crop / retouch / lens / drawn-mask). Cap 200 targets per call. Atomic. - ✅ wb_from_gray_card — derive temperature coefficients from a gray-card region. Closes survey Gap #20. - ✅ Vocab-load dtstyle-modversion drift check — sister to ADR-082's manifest-drift check; catches the bug class where a dtstyle's <module> byte disagrees with the engine pin (would have caught the v1.10.0-author-time bug that hung darktable for 60s × 25 entries).

ADR-088 / 089 / 090 ship as Draft; flip to Accepted on darkroom-session validation. The architectural decisions are locked; only visual-quality validation gates the status flip.

Closed / drafted in v1.9.0: - ✅ Real-raw visual-proof fixture path — closed via #103 (commit 830f89f); script extension supports HSL skip-listed entries against a real raw fixture; awaits the iguana fixture file drop. - ✅ Cinematic L2 looks — closed via #104; 9 named photographic recipes (cinematic / portrait / decade) compose existing primitives. - ✅ Gap-log analytics CLI — closed via #106; Phase 2 use-driven feedback loop now has read-side tooling. - ✅ CLI ergonomics polish — closed via #107; did-you-mean suggestions for unknown vocabulary names; shell-completion guide. - ✅ Session-log analytics CLI — closed via #109; sister to gap-log for session transcripts. - ✅ More L3 discrete kinds — closed via #110; 7 new clarity / sharpen / vignette / split-grade variants. - ✅ Lightroom-to-chemigram comparison guide — closed via #111; onboarding doc for Lightroom users. - ✅ Range masks (color-range / luminance-range) — closed via RFC-024 / ADR-085. Native parametric encoding via blendif: range_filter mask_spec field with kind ∈ {luminance, color_h, color_s, color_l}, AND-composes with drawn masks. 13 unit tests + 3 e2e tests against real darktable. Depth-range and pixel-perfect subject masks deferred to RFC-030 (deployed sibling-provider scaffolding); coarse subject identification works today via RFC-026 LLM-vision. - ✅ Spot removal / heal — closed via RFC-025 / ADR-087. New apply_spot MCP tool (sister to apply_primitive) with HEAL + CLONE algorithms on CIRCLE geometry. Wire-verified e2e against darktable. AI auto-detection ("find all the spots") deferred to RFC-030.

Still open (post-v1.9.0 horizon): - Manual tone curve (tonecurve) — the last Lightroom daily-use gap. Decoder is straightforward but the 520-byte spline-curve struct needs an empirical baseline from a darktable-GUI session. Tracked as #94, sequenced under #100. - Discrete colorzones HSL precision fallback — for the 5% HSL workflow that needs Lightroom's per-zone Range slider precision. Discrete-only; tracked as #98, sequenced under #100. - ✅ Coarse AI / content-aware masks — closed via RFC-026 / ADR-086. LLM-vision-as-provider: chat-client (Claude.ai / ChatGPT / Claude Code) sees the photo via render_preview and constructs mask_spec from spatial reasoning. Zero deployment cost; covers ~70% of content-derived masking workflows. - Precision-tier AI masks + auto-spot-detection — RFC-030 drafted (deferred). Deployed sibling-provider scaffolding for SAM-class subject masks, MiDaS-class depth, content-aware spot detectors. Unfreezes when LLM-vision precision becomes the bottleneck on real workflows. - Multi-photographer review phase plan — the deferred work from ADR-081's promotion threshold. Solo build-baseline → community transition. No RFC drafted yet. - Pack management / vendor packs / multi-pack composition stress-testing — the loader supports multi-pack but conflict resolution between packs hasn't been stress-tested. No RFC drafted yet.

Tier classification post-v1.9.0 in-progress: ADR-081's Tier 3 examples list is now stale (all of lens, denoiseprofile, colorzones-via-colorequal have been promoted or addressed). ADR-083 records the first formal Tier 3 → Tier 2 promotion (HSL via colorequal). ADR-085 promotes color/luminance range masks to Tier 2; ADR-087 promotes retouch (heal/clone) to Tier 2. The remaining named Tier 3 items (tonecurve, colorzones-as-discrete, AI-auto-spot-detection) are blocked on either darktable-session work or RFC-030's deployed-provider scaffolding.

What's been shipped against this survey's vision (v1.6.0 → v1.7.0)

This survey was the source document for RFC-021 (parameterization architecture) and RFC-022 (tiered baseline policy). Both closed; their ship landed across v1.6.0 and v1.7.0. The capability surface is materially different from when this doc was first drafted.

Architecture work completed:

RFC / ADR Decision Where it shows up
RFC-021 → ADR-077..080 Path C (decode/edit/re-encode op_params) is the default for explicitly-declared parameterizable modules All 18 parameterized entries below ride this
RFC-022 → ADR-081 Four-tier classification: Tier 0 immutable / Tier 1 Phase-4 floor / Tier 2 active expansion / Tier 3 default-opaque This document's "thin today" lists are Tier 3 watchlist
RFC-007 → ADR-082 Modversion drift: warn-loud at vocab load, hard-fail at apply via PatchError, env-var-strict mode Runtime safety net for the 11 Path C decoders

Vocabulary work completed (18 parameterized entries across 11 modules):

# Module Entries Tier
1 exposure exposure (single-axis EV) 1 (Phase 4 floor)
2 vignette vignette (single-axis brightness) 1
3 colorbalancergb saturation_global, vibrance, chroma_global, hue_angle, plus 4 brilliance axes (#86) 1 + 2
4 sigmoid sigmoid_contrast (single-axis) 1
5 bilat bilat_clarity_strength (single-axis) 1
6 grain grain_strength (single-axis) 1
7 highlights highlights_clip_threshold (single-axis) 1
8 temperature temperature (multi-axis: red_coeff, blue_coeff) 1
9 crop crop (4-axis: cx/cy/cw/ch — first workflow primitive) 2
10 sharpen sharpen (single-axis) 2
11 toneequal toneequalizer (9-axis: noise..speculars) 2

Plus the B&W trio (3 channelmixerrgb mv3 entries, Tier 0 discrete) and the 3 L2 looks (portrait / landscape / vintage_film).

Vocabulary delta against this survey's original snapshot:

Metric Survey premise Post-v1.7.0 Movement
Total entries 39 42 +3 net (16 collapsed into parameterized; 19 added)
Parameterized entries 0 18 +18
Modules with at least one entry 5 12 +7 (added bilat, highlights, temperature, crop, sharpen, toneequal, channelmixerrgb)
L2 composite looks 1 4 +3
Mask-bound entries 4 4 no change

Where this survey was the source of next-batch decisions

The survey's §12 "biggest user-visible holes" list (8 items) was the implicit target of RFC-022's Tier 2 ship. Status of each, post-v1.7.0:

# Original gap Status
1 toneequal zone-based tone equalizer ✅ shipped (Tier 2; 9-axis parameterized)
2 sharpen / diffuse-or-sharpen / equalizer family ⚠️ partial — sharpen shipped; diffuse-or-sharpen + equalizer still missing
3 denoiseprofile / nlmeans denoise ❌ Tier 3 (per-camera config; ADR-081)
4 clipping / ashift / flip composition ⚠️ partial — crop (= clipping module) shipped; rotation/perspective ❌
5 lens correction ❌ Tier 3 (lensfun-coupled; ADR-081)
6 retouch / spots portrait healing ❌ — would need its own RFC for the stroke-based input shape
7 colorzones HSL selective color ❌ Tier 3 (24+ params; ADR-081)
8 tonecurve manual tone curve ❌ — natural Tier 2 candidate; same multi-node shape as toneequalizer

3 of 8 fully closed; 2 partially; 5 remain (3 of those structurally Tier 3, 2 unblocked Tier 2 candidates).

What's next (post-v1.7.0)

Three shapes of work make sense for v1.8.0+:

Shape A — Tier 2 expansion of unblocked candidates (judgment-driven feature commits per ADR-081; no per-module ADR overhead): - tonecurve (manual curve) — same shape as toneequalizer, ~half a day - diffuse-or-sharpen strength axis — extends the sharpening family - flip / ashift — small structs, scalar parameters; closes the rotation/perspective half of #4 above - More L2 looks (cinematic, film-stock, decade, mood) — pure composition; no Path C

Shape B — Tier 3 promotion ADRs (each ships its own ADR per ADR-081; the bar is "real session evidence" + a workable cost-shape argument): - denoiseprofile — possible promotion shape: ship the camera's auto-detected profile baked in, expose only the strength axis - lens — possible promotion shape: L1 binding entry that auto-applies the profile from EXIF, no Path C decoding - colorzones — possible promotion shape: ship "axis-collapsed" entries (only the saturation curve at 4 nodes) rather than the full 24-param spline

Shape C — Novel-shape RFCs (architectural questions whose shape isn't yet settled): - retouch / spots — stroke-based input doesn't fit the per-image vocabulary entry pattern. RFC needed for what shape this takes. - AI / content-aware masks — explicitly conditional Phase 4 in IMPLEMENTATION.md; sibling project pattern (chemigram-masker-sam); biggest scope jump in the project.

What "growing it" actually requires

Per the project's Phase 2 framing: open darktable, capture moves you reach for that don't exist, drop the resulting .dtstyle into ~/.chemigram/vocabulary/personal/layers/L3/<module>/, and add a manifest entry. The vocabulary-authoring workflow is documented in docs/guides/authoring-vocabulary-entries.md. Building a personal pack to ~30–60 entries over 3 months is the design target.

What changed post-v1.7.0: most "growing it" work for parameterizable modules now ships as Path C decoders programmatically (no GUI seed needed — see docs/guides/expressive-baseline-authoring.md for the methodology). Hand-authoring via darktable GUI is reserved for entries whose photographic intent is genuinely discrete (Tier 0) — the 14 plain-discrete entries in expressive-baseline plus the starter-pack teaching artifacts.

11. The discrete-vocabulary problem (resolved for all 8 magnitude-ladder modules in v1.6.0+)

Status update — 2026-05-07: RFC-021 / ADR-077..080 closed this. All 8 magnitude-ladder modules from the Phase 4 plan are now parameterized: exposure, vignette, saturation_global, sigmoid_contrast, bilat_clarity_strength, grain_strength, highlights_clip_threshold, temperature (the first multi-parameter ship). The framing below is preserved as historical motivation. Whether to expand the parameterized baseline to net-new modules (sharpen / toneequal / denoise / lens / crop) is the open question in RFC-022.

The question (historical)

Why does chemigram ship expo_+0.3, expo_-0.3, expo_+0.5, expo_-0.5 as four separate vocabulary entries instead of one exposure(ev: float) primitive that takes a value at apply time?

Honest answer (historical)

The pre-v1.6.0 design was discrete named primitives — each entry was a fixed .dtstyle file with hardcoded parameter values. To get +0.7 EV you couldn't. You could apply expo_+0.5 and then expo_+0.3 separately (which worked because exposure stacks linearly), but that was a workaround, not a feature. If you wanted +1.5 EV, +2.0 EV, -1.0 EV, etc., you had nothing.

This was a real limitation. Combinatorially enumerating every plausible exposure value (every 0.1 EV from -3 to +3 = 60 entries) is infeasible. So is shipping only four.

How it's resolved (v1.6.0 onwards)

Per RFC-021 and the closing ADRs:

  • ADR-077 — Path C (decode/edit/re-encode op_params) is the default for explicitly-declared parameterizable modules. op_params opacity (ADR-008) still applies to non-parameterized modules.
  • ADR-078 — Vocabulary manifest gains a parameters array (multi-parameter from day one). Each parameter declares name / type / range / default plus a byte-level field location.
  • ADR-079 — apply_primitive accepts --value V (single-parameter shorthand) and --param NAME=V (repeatable, multi-parameter). MCP value arg is shape-polymorphic.
  • ADR-080 — Hard CI gate enforces 5-layer test coverage (unit / integration / lab-grade global / lab-grade masked / visual-proof sweep) on every parameterized entry.

The shipped surface today: chemigram apply-primitive --entry exposure --value 0.7 works. So does chemigram apply-primitive --entry vignette --value -0.6. Composes orthogonally with --mask-spec.

Why the project chose discrete-vocabulary in the first place

Three reasons, each defensible in isolation but they compound:

  1. Agent-paradigm fit. LLM agents reason better over named choices ("apply lift_shadows") than over continuous parameter tuning ("set exposure to +0.7"). The vocabulary becomes the action space; the agent picks moves the way a photographer does — by name and intent, not by knob position.

  2. op_params opacity (ADR-008). Every darktable module's parameters are stored as an opaque hex/base64 binary blob. To parameterize exposure(ev=0.7), the engine needs to decode the blob, edit the EV field, and re-encode it — a non-trivial reverse-engineering job per module ("Path C" in the project's terminology). The decision was to keep op_params opaque by default and only do Path C for high-value exceptions. Today, exposure is technically in the Path C exception list (per ADR-008's exception clause), but no parameterized API was ever built.

  3. Magnitude-as-vocabulary framing. The project's design system (concept/05-design-system.md) treats magnitude itself as a photographic decision — "subtle / medium / heavy" is the vocabulary, not a slider. This works for moves with discrete photographic intent ("vignette_subtle vs vignette_heavy"). It works less well for moves where the photographer's mental model is genuinely continuous ("I want +1 stop").

Where the framing breaks

The discrete-vocabulary framing makes sense for moves with photographic intent that fold cleanly into named buckets:

  • clarity_strong vs clarity_painterly — these are different kinds of clarity, not the same kind at different strengths. Names carry semantic weight.
  • grade_shadows_warm vs grade_highlights_warm — the zone matters; you can't slide between them.
  • The 4 mask-bound entries — each is a specific geometric move.

It breaks down for moves whose only meaningful difference is magnitude:

  • expo_+0.5 vs expo_+0.3 — these are not different photographic moves. They are the same move at two strengths. Shipping them as separate entries is a category error.
  • vignette_subtle vs vignette_medium vs vignette_heavy — three magnitudes of the same move.
  • wb_warm_subtle (no medium, no heavy) — the magnitude ladder is incomplete here.

The cleaner shape for this category: one entry per photographic intent, with a value parameter at apply time.

What this would look like

A parameterized apply-primitive flow:

chemigram apply-primitive iguana --entry exposure --value +0.7
chemigram apply-primitive iguana --entry vignette --value -0.4
chemigram apply-primitive iguana --entry wb_warmth --value +500   # Kelvin delta

The vocabulary entry declares which field of op_params is the magnitude field and what the valid range is. The engine decodes the blob, edits the field, re-encodes, applies. The agent still reasons in terms of named moves; magnitude becomes a parameter rather than part of the name.

This already has a precedent in darktable — every module's GUI is exactly this: named knob with a continuous value. Chemigram's discrete vocabulary diverges from darktable's underlying model.

What the migration would cost

Roughly, in scope-of-work terms:

  1. Path C decoders: per-module hex codecs for the modules where parameterization makes sense (exposure, vignette, temperature, colorbalancergb saturation, sigmoid contrast, grain strength, bilat clarity strength, highlights clip threshold). Probably 8–12 modules total. Each is a half-day of struct reverse-engineering against the darktable source. The _lab_grade_deltas.py test layer can validate the round-trip.

  2. Manifest schema extension: vocabulary entries gain an optional parameter field declaring the magnitude axis ({"name": "ev", "field_offset": 0, "type": "float", "range": [-3.0, 3.0], "default": 0.0}).

  3. CLI / MCP surface: apply-primitive gains a --value <n> flag (and MCP value arg). Engine routes through the parameterized synthesizer when the entry has a parameter field.

  4. Vocabulary cleanup: the 39 current entries collapse. expo_±0.3/±0.5/shadows_global_± become one parameterized exposure entry. The three vignette intensities become one. The two WB entries become one (with sign and magnitude as parameter). Net: probably ~20 parameterized entries replace ~30 discrete ones, plus space for ~10 new genuinely-discrete moves.

  5. ADR: a new ADR documenting the architectural shift. Likely supersedes part of ADR-008's Path A/Path C framing, since Path C becomes the default for parameterizable modules rather than the rare exception.

This is a real piece of architectural work — not a 1-hour ergonomic add. But it's also the right shape for what chemigram is trying to be. The current vocabulary's thinness on fundamentals is a symptom of the discrete-vocabulary framing being asked to do work it wasn't designed for.

Recommendation

Tomorrow's planning conversation should put parameterization on the table as a candidate for the next significant ship. It's the kind of change that unblocks the whole "fundamental controls" gap in section 1, 2, 3, 4, 6 above — without combinatorial vocabulary explosion.

12. Full darktable module catalog (the universe of what's possible)

What follows is every photographically-meaningful darktable 5.x module, organized by pipeline phase. The purpose is to make the gap between "what darktable can do" and "what chemigram exposes" explicit. A primitive is chemigram-shipped when there's at least one vocabulary entry touching the module today.

Categories: - ✅ shipped — chemigram has at least one vocab entry touching this module - ⚠️ partial — module is touched but coverage is thin (one direction, one magnitude, etc.) - ❌ not yet — module is not touched by any vocabulary entry - 🚫 out of scope — generally not exposed via vocabulary (camera/raw plumbing, output)

RAW phase (pre-demosaic / scene-data)

Module What it does Status
rawprepare Raw black / white levels, crop edges 🚫 out of scope (camera-specific)
temperature White balance (camera-as-shot or user) ✅ parameterized (Tier 1; multi-axis red_coeff + blue_coeff). Plus starter wb_warm_subtle discrete. Mask binding silently ignored (#80, documented limitation). Tint axis still missing.
highlights Highlight reconstruction (clip / LCH / inpaint / segmentation) ✅ parameterized (Tier 1; highlights_clip_threshold). Method enum preserved at default; method-selection variants would be Tier 0 discrete entries if needed.
hotpixels Hot-pixel removal 🚫 out of scope
rawdenoise Wavelet denoise on raw mosaic ❌ not yet
cacorrect Chromatic-aberration correction ❌ not yet
demosaic Bayer / X-Trans interpolation 🚫 out of scope (algorithm choice, not photographic intent)
denoiseprofile Profiled (per-camera) denoise ❌ not yet — high-value gap

Scene-referred phase (the modern pipeline middle)

Module What it does Status
exposure Global EV / black-level / clipping ✅ parameterized (Tier 1; range [-3.0, +3.0] EV). Plus 4 mask-bound variants (gradient × 2, ellipse, rectangle).
colorin Input color profile 🚫 out of scope (technical)
lens Lens correction (vignetting, distortion, CA) ❌ Tier 3 (lensfun-coupled per-camera config; ADR-081). Possible promotion via L1 EXIF-binding shape.
clipping Crop / aspect ratio / rotation ✅ parameterized as crop (Tier 2; 4-axis cx/cy/cw/ch). Aspect ratio constraint preserved at -1/-1 (free).
flip Orientation (rotate 90/180/270, flip) ❌ not yet — Tier 2 candidate (small struct, scalar parameter)
ashift Perspective / keystone correction ❌ not yet — Tier 2 candidate (small struct, scalar params)
liquify Local pixel pushing (manual warp) ❌ not yet — stroke-based input; would need novel RFC shape
retouch Frequency-separation retouching ❌ not yet — the major portrait gap; stroke-based input doesn't fit per-image vocabulary entry pattern. Would need its own RFC.
spots Spot healing / cloning ❌ not yet — same shape concern as retouch
censorize Pixelation / blur for privacy ❌ not yet
colorchecker Color calibration via reference chart 🚫 out of scope (technical)
channelmixerrgb RGB channel mixer (modern) ✅ 3 B&W entries (bw_convert, bw_sky_drama, bw_foliage); plus 4 parameterized colorbalancergb axes coexist via the colorbalancergb module — channelmixerrgb's other axes (RGB matrix, illuminant) remain Tier 3
colorbalancergb 4-way grade + saturation + chroma + brilliance ✅ shipped — 8 parameterized axes (saturation_global, vibrance, chroma_global, hue_angle, brilliance × 4) + 7 discrete (vibrance_+0.3 retired; per-zone chroma_boost × 3; grade × 4) = 15 entries on this module
tonecurve Manual RGB / Lab tone curve ❌ not yet — Tier 2 candidate; same multi-node-curve shape as toneequalizer
toneequal Zone-based tone equalizer (darktable's preferred local tone tool) ✅ shipped (Tier 2) — toneequalizer (9-axis: noise / ultra_deep_blacks / deep_blacks / blacks / shadows / midtones / highlights / whites / speculars; each ±2.0 EV)
sigmoid Modern tone mapping (filmic replacement) ✅ shipped — sigmoid_contrast (parameterized), plus blacks_lifted/crushed, whites_open (discrete kinds of tonal moves; Tier 0 by design). No whites_dampen ladder, no per-channel control.
filmicrgb Older tone mapping (still ships) 🚫 out of scope (sigmoid is preferred)
basicadj Combined exposure + contrast + saturation ❌ not yet (covered by separate primitives)
colorize Solid color tint ❌ not yet
colorbalance Older 4-way grade (pre-rgb) 🚫 out of scope (colorbalancergb is preferred)
colorzones Per-hue selective adjustment (HSL) ❌ Tier 3 (3 splines × 8 nodes = 24+ params; surface doesn't reduce cleanly to scalar --value V; ADR-081). Possible promotion shape: axis-collapsed entries (only the saturation curve at 4 nodes).
colorcontrast Lab a/b chroma contrast ❌ not yet
velvia Saturation with falloff for skin tones ❌ not yet
vibrance Old vibrance module (newer is in colorbalancergb) 🚫 out of scope (modern path)
monochrome B&W via filter color ❌ not yet — overlap with channelmixerrgb (B&W trio shipped there)
grain Film grain texture ✅ shipped — grain_strength (parameterized, range [0.0, 100.0]; Tier 1)
bilat (local contrast) Bilateral local contrast / clarity ✅ shipped — bilat_clarity_strength (parameterized strength axis; Tier 1) + clarity_painterly (Tier 0 — different kind, not strength)
bloom Soft glow / orton-style bloom ❌ not yet
soften Gaussian softening ❌ not yet
sharpen Unsharp mask sharpening ✅ shipped (Tier 2) — sharpen (single-axis amount; range [0.0, 2.0]). Radius + threshold preserved at darktable defaults.
diffuse (diffuse-or-sharpen) Reaction-diffusion sharpen / denoise / restore detail ❌ not yet — modern darktable's flagship sharpening; Tier 2 candidate (extends the sharpen family)
equalizer (contrast equalizer) Wavelet-based per-frequency contrast control ❌ not yet — Tier 2 candidate; per-frequency multi-axis surface
nlmeans Non-local-means denoise ❌ Tier 3 (denoise family; ADR-081)
bilateral (denoise) Bilateral denoise ❌ Tier 3 (denoise family)
defringe Color-fringe removal ❌ not yet
hazeremoval Atmospheric haze removal ❌ not yet
vignette Decorative radial darkening ✅ shipped (Tier 1) — vignette (parameterized; range [-1.0, +1.0]; negative darkens, positive lifts)
borders Frame border ❌ not yet (composition / output decoration)
watermark Text or image overlay ❌ not yet

Display-referred / output

Module What it does Status
colorout Output color profile 🚫 out of scope (technical)
gamma Display gamma 🚫 out of scope
dither Output dithering 🚫 out of scope
finalscale Resampling / output size 🚫 out of scope (handled by --width/--height)

Tally (post-v1.7.0)

  • darktable modules total: ~50 photographically-meaningful (excluding 🚫 out-of-scope plumbing)
  • chemigram ✅ shipped: 12 modules (exposure, sigmoid, colorbalancergb, vignette, grain, bilat, highlights, temperature, crop, sharpen, toneequal, channelmixerrgb)
  • chemigram ❌ not yet: ~28 modules — but the framing is sharper now: each unshipped module falls into one of three buckets per ADR-081's tiering policy and the survey's analysis above.

The remaining gaps, organized by feasibility (post-v1.7.0):

Unblocked Tier 2 candidates — the architecture supports these; they ship as feature commits when authored:

  1. tonecurve — manual tone curve. Same multi-node-curve shape as the already-shipped toneequalizer; ~half a day's work.
  2. diffuse-or-sharpen / equalizer — modern darktable sharpening family. sharpen already covers the unsharp-mask common case; these add reaction-diffusion + per-frequency control.
  3. flip / ashift — orientation + perspective correction. Small structs, scalar parameters; closes the rotation/perspective half of the composition gap.
  4. More L2 looks — cinematic / film-stock / decade / mood. Pure composition of existing primitives; same shape as the 4 looks already shipped.

Tier 3 candidates — default-opaque per ADR-081; promotion is evidence-driven and each gets its own ADR:

  1. denoiseprofile / nlmeans / bilateral — noise reduction family. Per-camera profiles violate Tier 2 cost-shape; possible promotion shape: ship the camera's auto-detected profile baked in via L1 EXIF binding, expose only the strength axis.
  2. lens — lens correction. Lensfun-coupled per-lens database lookup. Possible promotion shape: L1 binding entry that auto-applies the profile from EXIF (sidesteps the parameterized-vocabulary framing entirely).
  3. colorzones — HSL selective color. 3 splines × 8 nodes = 24+ params; doesn't reduce cleanly to --value V. Possible promotion shape: axis-collapsed entries (only the saturation curve at 4 nodes).

Novel-shape gaps — would need their own RFCs:

  1. retouch / spots — frequency-separation retouching + spot healing. The major portrait gap. Stroke-based input doesn't fit the per-image vocabulary entry pattern; the right shape isn't yet clear.
  2. AI / content-aware masks ("the manta's belly") — explicitly conditional Phase 4 in IMPLEMENTATION.md; sibling project (chemigram-masker-sam) shape; biggest scope jump in the project.

The gap isn't a single uniform "30 modules to author" anymore — it's three distinct kinds of work each with its own decision shape. Items 1–4 are routine Tier 2 expansion; items 5–7 need policy decisions before authoring; items 8–9 need design RFCs before policy.

13. Lightroom daily-use panel mapping

Comparison against the panels a typical Lightroom photographer touches on every photo (the user's own daily-use list, captured 2026-05-07). Two purposes: (1) help users coming from Lightroom understand what's available; (2) ground "what's missing" in the most-used controls of the dominant photo editor — a sharper signal than the abstract "modules darktable has" framing in § 12.

The 18 daily-use controls organized by Lightroom panel; mapped to chemigram capability with explicit status flags:

Light panel (7 controls)

Lightroom control chemigram equivalent Status
Exposure exposure --value V (range [-3.0, +3.0] EV) ✅
Contrast sigmoid_contrast --value V (range [0.5, 5.0]; 1.5 = no-op) ✅
Shadows (zone lift) toneequalizer --param shadows=V + gradient_bottom_lift_shadows (mask-bound) ✅
Blacks (deep-shadow point) toneequalizer --param deep_blacks=V / --param blacks=V; plus blacks_lifted / blacks_crushed (discrete kinds) ✅
Whites (white point) toneequalizer --param whites=V; plus whites_open (discrete) ✅
Highlights (highlight roll-off) highlights_clip_threshold --value V + toneequalizer --param highlights=V + gradient_top_dampen_highlights (mask-bound) ✅
Tone Curve (parametric / point curve) — ❌ — tonecurve module unrepresented; tracked as #94 (520-byte spline-curve struct, needs darktable GUI baseline capture). The 9-band toneequalizer covers most intents but isn't the same idiom.

Light: 6/7 fully covered, 1 missing (#94 tone curve).

Color panel (4 controls)

Lightroom control chemigram equivalent Status
WB Temperature (Kelvin slider) wb_kelvin_delta --param kelvin_delta=V (range [-3000, 3000]; positive = warmer) — shipped via #102 / commit <TBD> as a UX wrapper on the existing temperature decoder. The raw temperature --param red_coeff=V --param blue_coeff=V axes remain available for users who want direct coefficient control. ✅
WB Tint (green ↔ magenta slider) wb_kelvin_delta --param tint_delta=V (range [-200, 200]) shipped via #102; or temperature --param green_coeff=V for direct coefficient control (#90 Bucket A.3 / commit 1a00254) ✅
Vibrance vibrance --value V (range [-1.0, +1.0]) ✅
Saturation saturation_global --value V (range [-1.0, +1.0]) ✅

Color: 4/4 fully covered.

Color Mixer panel — HSL per color band (1 panel = 24 controls)

Lightroom exposes Hue + Saturation + Luminance per 8 color bands (red / orange / yellow / green / aqua / blue / purple / magenta) = 24 sliders. Shipped via RFC-023 / ADR-083 (commit 1b5db21) — backed by darktable's modern colorequal module (mv4, 128-byte flat struct), not the older colorzones spline-curve module.

Lightroom control chemigram equivalent Status
HSL Hue per color (8 axes) hsl_hue --param hue_<color>=V (8 axes: hue_red, hue_orange, hue_yellow, hue_green, hue_cyan, hue_blue, hue_lavender, hue_magenta; each [-180, 180] degrees) ✅
HSL Saturation per color (8 axes) hsl_saturation --param sat_<color>=V (8 axes; each [-1.0, 1.0]) ✅
HSL Luminance per color (8 axes) hsl_luminance --param bright_<color>=V (8 axes; each [-1.0, 1.0]) ✅

Color Mixer: 24/24 covered. Backed by colorequal mv4. The 5% spline-curve precision use case (Lightroom's HSL Range slider per-zone falloff) is tracked as #98 — discrete-only colorzones fallback if/when needed.

Color Grading panel (multi-axis)

Lightroom's color grading panel offers per-zone (shadows / midtones / highlights / global) hue + saturation wheels, plus per-zone luminance, plus global blending and balance. Bucket A.5 (#91 / commit 7eb4aab) added 9 axes here.

Lightroom control chemigram equivalent Status
Shadows: hue + saturation hue_shadows (parameterized [0, 360°]) + saturation_shadows (parameterized [-1, 1]) + discrete grade_shadows_warm / grade_shadows_cool ✅
Midtones: hue + saturation hue_midtones (parameterized [0, 360°]) + saturation_midtones (parameterized [-1, 1]) + discrete grade_midtones_warm / grade_midtones_cool (#90 Bucket A.4 / commit 72ff3e9) ✅
Highlights: hue + saturation hue_highlights (parameterized [0, 360°]) + saturation_highlights (parameterized [-1, 1]) + discrete grade_highlights_warm / grade_highlights_cool ✅
Global: hue rotation hue_angle --value V (range [-180.0, +180.0] degrees; rotates all pixels uniformly) ✅
Per-zone luminance brilliance_shadows, brilliance_midtones, brilliance_highlights, brilliance_global (all parameterized; range [-1.0, +1.0]) ✅
Blending (zone falloff) shadows_weight, highlights_weight (parameterized [0, 4]) ✅
Balance (shadow/highlight midpoint) white_fulcrum (parameterized [-2, 2]) ✅

Color Grading: 7/7 controls fully covered.

Effects panel (5 controls)

Lightroom control chemigram equivalent Status
Texture (mid-frequency detail) texture --param first=V --param second=V --param sharpness=V via diffuse-or-sharpen (#92 Bucket A.6 / commit c9dfe83). Algorithm match for Lightroom's mid-frequency Texture work; first is the primary axis. ✅
Clarity bilat_clarity_strength --value V (parameterized) + clarity_painterly (different kind of clarity — softer; Tier 0) ✅
Dehaze (atmospheric haze removal) dehaze --param strength=V --param distance=V via hazeremoval (#90 Bucket A.2 / commit 949516f; range [-1, 1] for strength, negative adds atmospheric fog) ✅
Vignette vignette --value V (range [-1.0, +1.0]) ✅
Grain grain_strength --value V (range [0.0, 100.0]) ✅

Effects: 5/5 covered.

Transform panel (5 axes)

Lightroom's Transform panel: rotation, vertical/horizontal perspective, shear, aspect adjust. Architectural / interior / flat-art photographers reach for this every shoot. Bucket A.7 (#101 / commit <TBD>) ships via darktable's ashift module.

Lightroom control chemigram equivalent Status
Rotate (degrees) transform --param transform_rotation=V (range [-180, 180]°) ✅
Vertical perspective (keystone) transform --param transform_lensshift_v=V (range [-1, 1]) ✅
Horizontal perspective transform --param transform_lensshift_h=V (range [-1, 1]) ✅
Shear transform --param transform_shear=V (range [-1, 1]) ✅
Aspect adjust transform --param transform_aspect=V (range [0.5, 2.0]; default 1.0) ✅

Transform: 5/5 covered.

Lens-tuning fields (focal length, crop factor, ortho-correction) and the user-drawn-lines reference storage are preserved verbatim — those are darktable-GUI-authored when the photographer wants line-based perspective correction; the transform entry covers the slider-based daily-use surface.

Daily-use summary

Aggregating across the 5 panels — Color Mixer's 24 sliders count as 24 controls now that they're individually addressable, Color Grading's per-zone H/S/L axes count individually, and Transform is the final added panel post-#101:

Panel Controls ✅ Full ⚠️ Partial ❌ Missing
Light 7 6 0 1 (#94 tone curve)
Color 4 4 0 0
Color Mixer 24 24 0 0
Color Grading 7 axes 7 0 0
Effects 5 5 0 0
Transform 1 panel (5 axes) 5 0 0
Total daily-use surface 52 distinct controls 51 0 1

Lightroom daily-use parity: 51/52 fully shipped (98%) + 1 deferred (#94 tone curve, blocked on darktable-GUI baseline session per #100 umbrella).

What this implies for next work

Daily-use Lightroom parity is essentially closed. The only remaining gap is #94 tone curve, which is blocked on the darktable-session empirical-baseline work tracked under #100. Everything else in the daily-use surface ships.

Post-v1.9.0-in-progress, the next work splits into darktable-bound and keyboard-only tracks:

Darktable-bound (the #100 umbrella):

  1. #94 manual tone curve decoder — needs darktable-GUI baseline capture
  2. #95 lens EXIF auto-binding — needs real-raw + lensfun lookup verification
  3. #96 denoise wavelet-curve baseline verification — confirm or fix the constructed x[6][7]/y[6][7] baseline
  4. #98 colorzones spline-curve HSL precision fallback — discrete-only presets
  5. HSL real-raw visual proofs — the iguana fixture drop closes the v1.8.0 #103 placeholder rows

Keyboard-only (no darktable needed):

✅ Closed during v1.9.0: - RFC-024 / ADR-085 (range masks), RFC-026 / ADR-086 (LLM-vision masks), RFC-029 / ADR-084 (compositional masks), RFC-025 / ADR-087 (spot removal) - chemigram vocab validate <name> CLI verb - chemigram cache {list, size, clear} CLI sub-app - 5 compositional-mask L2 looks demonstrating the full mask trilogy

📝 Drafted during v1.9.0 (deferred): - RFC-030 — deployed sibling-provider scaffolding (precision-tier AI subject + depth + auto-spot-detection)

Still open:

  1. RFC-027 — Multi-photographer review phase plan — the deferred work from ADR-081's promotion threshold. Solo build-baseline → community transition. The Tier 3 → Tier 2 evidence threshold, gap-log methodology, vocabulary conflict resolution between photographers, the social shape of pack contribution.
  2. RFC-028 — Pack management / multi-pack composition / vendor packs — the loader supports multi-pack but conflict resolution between packs hasn't been stress-tested. Argues the architectural shape for vendor-distributed packs and pack-versioning.
  3. Recipe book / cookbook — "100 ways to use chemigram." Worked photographic recipes pulling from the 81-entry vocabulary corpus + the new mask trilogy.
  4. Onboarding guide for new contributors — what a new contributor reads first (concept → architecture → CLAUDE.md → recipe book).
  5. Architecture diagrams — visual one-pager(s) of the stack (agent → MCP/CLI → engine → darktable).
  6. Property-based fuzz tests for parameterize decoders — fuzz the 18 patch() functions + the new parametric mask + retouch encoders to surface byte-correctness regressions. Catches issues the 5-layer manual tests don't.

The --without-darktable track is still rich; the with-darktable track is bounded by the #100 umbrella. See § 13's "What this implies" for the Lightroom-parity slice specifically.

See also