1.0.0-alpha.2
Gerstep/cybos1.0.0-alpha.2Jul 27, 2026by tyt2y3
AI Summary
Major update introducing watershed clustering, curve simplification, and adaptive thresholding, along with a significant refactoring of the CLI and API to use a `clustering` selector.
Key Highlights
- New watershed clustering algorithm with interactive detail slider
- Curve simplification (`--simplify`) to reduce file size
- Binary thresholding with adaptive lighting support
- Renamed `color_mode` to `clustering` across CLI and bindings
Breaking Changes
- `color_mode` replaced by `clustering` (color-cluster | bw | watershed)
- Spline tuning flags (`--corner-threshold`, etc.) are hidden from CLI
New Features
- Hierarchical watershed region forming frontend
- Curve simplification using Schneider re-fit
- Binary and adaptive thresholding options
- Cutout mode merging nearby faces
- Fixed spline fitting for thin strands
Full Release Notes
### Added * Watershed clustering (`--clustering watershed`): a new region-forming frontend — a hierarchical watershed on the pixel graph (Cousty et al. 2009; Najman et al. 2013), controlled by one dial, `--watershed-detail` (0..=255; each +25.5 roughly doubles the region count). Regions follow image content, with no watershed-line pixels. * Boundaries come out calm: antialiased pixels snap to the color-midpoint iso-line instead of meandering with the noise inside the ramp. * `stacked` stacks the merge tree itself (coarse ancestors below, refined regions on top), so overdraw stays seam-free. * `cutout` gets the partition natively; neighbouring faces closer than `max(2, (255 − detail) / 8)` merge, so faces a human cannot tell apart never survive as separate patches. * `WatershedHierarchy` is public, split into `build` (expensive, image-only) and `cut` (near-instant); `Session` re-cuts a cached hierarchy on detail changes, making the slider fully interactive (~25 ms vs ~40 ms on a 1400×775 photo). * Curve simplification (`--simplify <tolerance>`, `Config::simplify`, `simplify` in Python and Node; off by default): a paper.js-style Schneider re-fit — each smooth run between corners is redrawn with the fewest cubics that stay within the tolerance (px). Roughly halves file size (sample photo at tolerance 1: 229 → 138 KB stacked, 103 → 36 KB watershed cutout). Runs on fitted geometry before composition, so cutout simplifies each shared boundary once and stays seam-free; corners and junction endpoints stay pinned. * Binary thresholding: a tunable fixed threshold (`--threshold`) and Bradley–Roth adaptive thresholding for uneven lighting (`--adaptive`, `--adaptive-window`, `--adaptive-t`) — also on `Config`, Python, and Node. * Cutout mode merges neighbouring faces whose colors are within one gradient step, rejoining the near-identical faces that stacked gradient layering splits a smooth area into. ### Changed * `color_mode` is replaced by `clustering` (`color-cluster` | `bw` | `watershed`) across the CLI, Rust, Python, and Node — the field selects the region-forming algorithm, not a color space. * The spline fine-tuning flags (`--corner-threshold`, `--segment-length`, `--splice-threshold`) are hidden from CLI help — still accepted, but without their `-c`/`-l`/`-s` short forms. The defaults serve virtually every conversion; `--simplify` supersedes them. ### Fixed * Spline fitting no longer swings far away from the outline around thin strands (a long-standing defect, fixed via visioncortex 0.9.1): a sparse splice slice could be fitted by a single cubic that passed through every sample yet ballooned up to ~30 px sideways between them. Slices are now densified before fitting and multi-cubic fits kept in full, in both stacked mode and the mosaic fitter.