# craft-develop The raw develop pipeline, CPU path: the shared pipeline of [#587](https://github.com/storytold/craft-libs/issues/587), moved from LightCraft one module at a time. No `unsafe`, no panics on hostile input, no file I/O, and it builds for `wasm32-unknown-unknown`. ```toml craft-develop = { git = "https://github.com/storytold/craft-libs", rev = "" } ``` ## What is here | Module | From LightCraft (`a75497d6`) | Contents | |---|---|---| | `color::spline` | `crates/color/src/spline.rs` | Re-exports `craft_color_transform::curve` (`MonotoneCurve`, `Lut1`). | | `curve` | `crates/pipeline` | The parametric tone curve and its regions. | | `finish` | `crates/pipeline` | Pure-Rust finishing helpers: sharpening masks/limits, vignette parameters, value grain, gamut mapping, and curve saturation refinement. | | `lut` | `crates/pipeline` | 3D lookup tables and the `.cube` reader. | | `tone` | `crates/pipeline` | Global tone mapping (`ToneMap` and `CameraTone`) from scene luminance to display-linear luminance, SDR and HDR. | | `inpaint` | `crates/inpaint` | Content-aware hole filling for spot removal: `complete` (multi-scale patch completion with PatchMatch) and `membrane_fill` (a smooth harmonic fill, its seed and fallback). Interleaved `f32` buffers of any channel count, a fixed-seed RNG, the same result on any number of threads. | ## Parity with LightCraft Each module comes with a test that its output is LightCraft's at the base commit, compared with [`craft-render-parity`](../render-parity/README.md) at `Tolerance::BASELINE`. The expected pictures in `tests/fixtures/` were written by LightCraft's own crate at that commit, on the scene the test builds. Differences from LightCraft (ordinary-range output is covered by the baseline fixtures): - `inpaint::complete_with` returns `None` for a patch radius wider than the image before computing with it (LightCraft reaches the same `None`, or overflows on a huge radius), returns an image with no pixels at once, and reports `Cancelled` when any one poll said so, including a poll made on a worker thread. - `membrane_fill` scales extreme finite boundaries for the harmonic solve and bounds the result by those boundaries, avoiding overflow in means and neighbour sums. Known pixels remain unchanged; ordinary values take the original arithmetic path. This does not promise arbitrary extreme-value safety for PatchMatch completion, which also computes squared costs and downsamples its input. - Membrane cell indices retain `usize` width. PatchMatch rejects grids whose signed search offsets, summed-area counts or padded table cannot be represented; patch-area arithmetic no longer overflows `i32`. An already-cancelled oversized-patch request returns `Cancelled`, not an unavailable-patch result. ## Not here yet The rest of the pipeline: the rest of colour, raster, the sensor stage, raw loading and the render stages. No app uses this crate yet.