π Geometric Computer Vision Library for Spatial AI
Python artificial-intelligencecomputer-visiondeep-learninghacktoberfestimage-processingmachine-learningneural-networkpythonpytorchroboticsspatial-ai
chore: start 0.9.1 development after verified 0.9.0 release
Dmytro Mishkin6d4dada5acf3
diff --git a/ROADMAP.md b/ROADMAP.mdindex 169634998..4954dcc5d 100644--- a/ROADMAP.md+++ b/ROADMAP.md@@ -20,6 +20,20 @@ measured in confidence, not API count β a strong release can contain zero newAPIs and still deliver better `torch.compile` coverage, fixed convention ambiguity,improved numerical stability, and stronger dtype/device behavior.+## Release status β 2026-10-23++[Kornia 0.9.0](https://github.com/kornia/kornia/releases/tag/v0.9.0) is released and+verified. The published PyPI wheel and source archive match the CI build and GitHub+assets. Fresh Python 3.11/3.13 installs passed imports and image I/O; the installed+wheel passed the CPU/MPS workflow smoke checks, including compiled RANSAC on CPU.+Stable and versioned documentation serve the 0.9.0 LLM files.++Development now targets `0.9.1.dev`. The next patch prioritizes confirmed regressions+and compatible fixes. Box/camera/augmentation convention decisions, the external+conformance corpus, and the proposed Tier A policy remain separate longer-term+work under the existing stability policy. The current release sequence and+follow-ups are tracked in [#4620](https://github.com/kornia/kornia/issues/4620).+## Guiding themesFour engineering north stars shape the roadmap below.@@ -111,22 +125,25 @@ maintainer with model-work capacity joins the project. Contributions already inflight before the freeze (an open issue plus a PR under review) are grandfatheredand finish under the rules they started under.-## Short term β next release--- **Stabilize CI and the augmentation core.** Recent fixes repaired regressions from the- ONNX-exportability refactor and healed the scheduled test matrix. Ship these in a- release so users get the fixes.-- **Convention-block batches.** Continue the per-op audit:- `geometry.conversions` β bounding boxes- ([#3934](https://github.com/kornia/kornia/issues/3934)) β camera β augmentation- classes. Each batch delivers Convention blocks, pinning tests, and dedicated- issues for any behavior bug found.-- **Broken-output bugfixes** from the audit, landing ahead of the repair window:- pyramid bounds check ([#3927](https://github.com/kornia/kornia/issues/3927)),- 1-pixel crop NaN ([#3929](https://github.com/kornia/kornia/issues/3929)).-- **CI truthfulness.** Every advertised CI matrix axis must actually reach the code- under test; first repair is the unused pytorch-version axis- ([#3930](https://github.com/kornia/kornia/issues/3930)).+## Short term β 0.9.1 maintenance candidates++These are priorities to investigate and validate, not a promise that every item+will ship in 0.9.1. Confirmed downstream regressions take precedence.++- **Bilateral-filter half gradients:** restore accuracy for learnable tensor sigmas+ without losing backend performance ([#6284](https://github.com/kornia/kornia/issues/6284)).+ Float32 sigmas remain the documented workaround until a fix is verified.+- **Compiled RANSAC cache safety:** isolate the CPU-to-CUDA cache collision and+ verify a fix or mitigation ([#6278](https://github.com/kornia/kornia/issues/6278)).+ Separate `TORCHINDUCTOR_CACHE_DIR` directories per device, or `KORNIA_RANSAC_AOT=0`,+ remain the current workarounds.+- **CUDA reliability:** triage the Kannala-Brandt TorchScript second-call hang+ ([#6257](https://github.com/kornia/kornia/issues/6257)); consolidate the 29 recorded+ pre-existing CUDA-half test failures into explicit tracking and coverage work.+ CUDA-half support is not a blanket patch-release guarantee.+- **Post-release checks and documentation:** confirm the next scheduled cross-platform+ PyPI-package test, retain the verified release evidence, and decide how stable/versioned+ LLM documents should link to their matching documentation rather than `/en/latest/`.- **Contribution documentation:** keep the social contract, technical guidance, andproject automation consistent and easy to understand.- **Progressively enable more `ruff` rule sets** to raise code healthdiff --git a/kornia/__init__.py b/kornia/__init__.pyindex 42d6b7390..c02a3a6fc 100644--- a/kornia/__init__.py+++ b/kornia/__init__.py@@ -100,4 +100,4 @@ def __getattr__(name: str) -> Any:# Version variable-__version__ = "0.9.0"+__version__ = "0.9.1.dev"