Skip to content

ADR-094 — Camera-aware denoiseprofile threshold scaling

Status · Draft (impl shipped; flips to Accepted on darkroom validation) Date · 2026-05-23 TA anchor · /components/synthesizer · /constraints/opaque-hex-blobs Related RFC · RFC-039 (follow-up sibling under ADR-091 machinery)

Context

RFC-039's engine machinery (ADR-091) ships raw_path threading through apply_entry and per-module declaration of raw-metadata needs. ADR-093 used the machinery for the first raw-derived parameter — camera-aware WB on the temperature module. The RFC explicitly named denoise as the next module to fold into the same pattern: a denoise threshold authored against ISO 200 should scale up automatically when applied to an ISO 6400 raw, because shot noise grows roughly as √ISO.

Before this ADR, chemigram.core.parameterize.denoiseprofile.patch() operated in raw bytes — the photographer authored a fixed threshold against a specific shot's ISO and the same bytes applied unchanged to every other shot. A daylight-ISO threshold over-denoised low-ISO output (smearing detail) and under-denoised high-ISO output (visible chroma noise).

Decision

chemigram.core.parameterize.denoiseprofile.patch() gains an optional raw_path: Path | None keyword argument. When supplied:

  • The function reads the raw's ISO via chemigram.core.exif.read_camera_iso(raw_path) (rawpy's other.iso_speed, with exifread fallback).
  • The source threshold values in the denoiseprofile op_params are scaled by 2 ** log2(iso / _DENOISE_REFERENCE_ISO) where _DENOISE_REFERENCE_ISO = 200. The factor is 1.0 at ISO 200, 2.0 at ISO 400, 4.0 at ISO 800, 8.0 at ISO 1600, etc.
  • The scaling applies multiplicatively to the strength axis; the chroma/luma split and frequency profile stay at the source values.
  • Explicit strength overrides bypass the substitution — the caller is opting into raw-strength control.

If raw_path is None (synthetic chart fixtures), behavior is unchanged. If the raw is unreadable, the substitution silently falls back to source.

Rationale

Shot noise in a sensor's raw output grows approximately as √ISO, so a noise-suppression threshold tuned for ISO 200 needs to grow roughly linearly with ISO to maintain the same perceptual smoothness. The log2-scaled factor above ISO 200 (the reference) tracks the ISO doubling cadence: every ISO doubling doubles the threshold, matching the noise floor's doubling. Below ISO 200 (clean territory) the factor stays at 1.0 — over-attenuation here would smear detail that doesn't need denoising.

Reading ISO via rawpy is the architecturally consistent path with ADR-093's WB substitution: same engine API, same lazy reader, same raw_path parameter, same fallback semantics. The vocabulary author writes one entry; the entry adapts to whatever the photographer's raw was shot at.

Alternatives considered

  • Linear ISO scaling (factor = iso / 200). Rejected because shot noise grows as √ISO, not linearly — linear scaling would over-attenuate high-ISO output. Log2 sits between √ISO (1.41× per stop) and linear (2× per stop) but matches the ISO-doubling cadence cleanly.
  • Strength curves per camera body. Rejected because the per-body strength tables would require empirical calibration the project doesn't have bandwidth for. Log2 scaling is body-agnostic and matches the photographic intuition "two stops of ISO = two stops more denoise."
  • Skip the substitution at low ISO (e.g., < 200). Considered and adopted by-omission: the formula already returns 1.0 at the reference ISO and below; no explicit branch needed.

Consequences

Positive: - Denoise vocabulary entries port across ISO without per-shot reauthoring. - Same engine machinery as ADR-091 / ADR-093 — no new infrastructure. - Photographer's intent (e.g., "remove visible noise") survives ISO changes.

Negative: - The log2 formula is a heuristic, not a calibrated noise model. At extreme ISOs (above ISO 25600) the formula may over-attenuate; per-body calibration would be more accurate. - ISO probes via rawpy's other.iso_speed can be None for some bodies; the fallback is "use source threshold" which is the safe degradation.

Implementation notes

  • src/chemigram/core/parameterize/denoiseprofile.py — patch() accepts raw_path kwarg
  • src/chemigram/core/exif.py — read_camera_iso(path) helper
  • tests/unit/core/parameterize/test_denoiseprofile.py — ISO scaling tests
  • Closes RFC-039 follow-up question 2 (denoise per ISO scaling).