v0.5.0

Highlights

• Added a public C API and libdinov2 target, with opaque model/context handles, batch encoding, and model loading from files, buffers, or callbacks.
• Added a WebAssembly SIMD128 build and browser demo, now hosted at https://espetro.github.io/dinov2.cpp/.
• Added an optional HTTP embeddings server, plus deduplication, container, and CI visual-regression examples. These remain tier-2, best-effort surfaces.
• Fixed inherited inference and loader issues, including flash-attention padding corruption, GGUF validation, backend device-pointer reads, model-load memory leaks, and positional embedding interpolation.
• Added backbone-only GGUF conversion and publishing for all eight DINOv2 variants, alongside the eight classifier variants. Backbone models support feature extraction and reject classification.
• Added ARM64 benchmark evidence, nightly parity checks, and the JSONL output contract check.
• Improved large-image safety, preprocessing modes, batching, binary embeddings metadata, and clean allocation errors from v0.4.0.
• Added Joaquin Terrasa's copyright notice while retaining the original upstream copyright notice. The project remains MIT licensed.

Compatibility and limitations

• The C API is still unstable during the 0.x series. Pin a release when embedding it.
• The browser demo uses single-threaded WebAssembly SIMD128. Native CPU performance is faster.
• The server and examples are tier-2 and are not production-service guarantees.
• DINOv2 is image-only, not a CLIP-style text-image model. Image decoding is limited to the formats supported by stb_image.
• Benchmark results are hardware-specific; consult docs/benchmarks.md and docs/parity/ for methodology and measurements.

Assets

Release binaries are attached below with SHA256 checksums. GGUF model weights are published separately under https://huggingface.co/dinov2-cpp-core.

v0.4.0

Highlights

• Embeddings JSON/JSONL records with a stable contract (index, image, grid, row-major patches), plus a D2EMB v2 preview binary format carrying grid dims.
• Batched and multi-image inference via --batch and repeated -i inputs.
• Feature-mode preprocessing is now bounded by default (--preprocess bounded, shortest edge 518 + true-ceil patch alignment), with hf (HF AutoImageProcessor recipe) and crop518 (fixed 37x37 grid) modes, --no-resize opt-out, and a --max-tokens cap enforced before graph construction.
• Strict numeric parsing for all value flags; graph and model buffer allocation failures exit cleanly instead of aborting.
• Backbone-only DINOv2 GGUF loading and conversion support.
• Parity and benchmark evidence for both small checkpoints (8/8 checks each), plus register-token guidance for feature workflows.

Changelog

[0.4.0] - 2026-09-21

Added

• Return cls, pooled, and score embeddings from dino_predict
• Add output-mode, version, and help flags to CLI parsing
• Emit embeddings JSON and restore top-k output in CLI
• (scripts) Add parity check vs PyTorch reference
• (params) Add n_batch and --batch flag with bounds checking
• (graph) Batch-aware encoder graph in forward_features and attn
• (predict) Multi-image dino_predict API
• (cli) Accept multiple -i inputs and comma-separated image lists
• (cli) Run multi-image inference in n_batch chunks
• (cli) Report per-image throughput in bench output
• (cli) Add preview binary embeddings output
• Harden backbone-only DINOv2 loading
• (preprocess) Bound feature mode with --preprocess, --no-resize, --max-tokens
• (cli) Add index and grid to records, D2EMB v2 header

Fixed

• Preserve aspect ratio in classify preprocess resize
• Pool over actual patch-token count in classify head
• Route dino_model_load logs to stderr
• (scripts) Gate parity on patch flat+mean cosine, keep token min informational
• (dinov2) Exclude register tokens from classify pooling
• (dinov2) Bounds-check value flags in dino_params_parse
• (scripts) Fail parity check on empty image list and CLI timeout
• Add missing scale param to attn declaration
• (cli) Address batched inference review findings
• (cli) Enforce strict numeric parsing
• (cli) Address Task 1 review nits
• (cli) Address binary embeddings review findings
• (ci) Use supported Hugging Face CLI
• (bench) Parse benchmark JSON portably
• Address final parity review findings
• Support backbone-only DINOv2 conversion
• (converter) Harden DINOv2 backbone resolution
• Check ggml graph and model buffer allocation failures

Changed

• (readme) Drop Topics section and topic-count badge
• (plans) Point CUDA smoke-test step at the live HF slug
• Apply clang-format-18 to bench path + bench flag declarations
• Drop unused print_t_f32 debug helper
• Cover classify preprocess aspect and l2_normalize
• Ignore CMake preset build-* directories
• Add cli.md reference and sync CLI docs with embeddings output
• Complete --help blocks and fix model path in cli.md example
• (cli) Fix build preset, threads default, and flag-table lead-in
• Update agent memory for embeddings-dropin learnings
• Batch inference coverage with synthetic tiny model
• Cover flash-attention batch path in dino_predict
• Document batch inference across CLI reference and guides
• (spec) Define register-token recommendations
• Recommend register-token models for feature workflows
• Specify parity benchmark and binary preview
• Add reproducible parity evidence
• (parity) Clarify evidence provenance and downloads
• (bench) Report blocked Ubuntu evidence run
• Record Ubuntu benchmark rerun blocker
• (bench) Record Ubuntu benchmark artifact
• (bench) Preserve model provenance
• (bench) Record current Ubuntu benchmark evidence
• Harden backbone loader review coverage
• (bench) Publish Ubuntu-only benchmark evidence
• Refresh final parity evidence
• Refresh final parity evidence
• Fix regular GGUF checksum
• (spec) Design large-image preprocessing bound and record contract
• Bump project version to 0.4.0
• (spec) Settle preprocess modes, true-ceil alignment, and record contract
• (preprocess) Single resample to bounded target dims, true-ceil ctx sizing
• Record contract, preprocess modes, D2EMB v2
• (parity) Record measured hf-mode residual for tench
• (parity) Record rerun code state and CLI 0.4.0

Miscellaneous

• Add project-scoped mise.toml for clang-format-18 + git-cliff
• Define DINOV2_VERSION from project version
• Add cmake buildPresets so --preset build works
• Merge pull request #10 from espetro/feat/embeddings-dropin
• Merge pull request #11 from espetro/feat/batched-inference
• Merge pull request #12 from espetro/feat/batched-inference
• Merge pull request #13 from espetro/feat/batched-inference
• Merge pull request #14 from espetro/fix/large-image-preprocessing

Full Changelog: v0.3.0...v0.4.0

v0.3.0

Full Changelog: v0.2.0...v0.3.0

v0.2.0

What's Changed

• feat: ggml v0.24.0, HF GGUF publishing pipeline, README rewrite by @espetro in #2
• ci(refactor): port gguf publish to bash script + use ubuntu-latest-large runner by @espetro in #3
• ci(cache): cache .venv-publish + document HF_TOKEN by @espetro in #4
• ci(gguf): fit conversion within ubuntu-latest (~14 GB) via lean deps + cache cleanup by @espetro in #5
• ci(gguf): hardcode HF_HOME (runner.home not allowed in env:) by @espetro in #6
• perf(gguf): relocate caches to /mnt, hf_xet uploader, dynamic matrix, per-run audit logs by @espetro in #7
• fix(ci): emit JSON array from plan job (fromJson requires it, not a space-separated string) by @espetro in #8

Full Changelog: v0.1.0...v0.2.0

v0.1.0

What's Changed

• feat: v0.1.0 - drop OpenCV, cross-platform CI + releases, dev harness by @espetro in #1

New Contributors

• @espetro made their first contribution in #1

Full Changelog: https://github.com/espetro/dinov2.cpp/commits/v0.1.0