A small, MIT-licensed, agent-first orchestration CLI, implemented in Rust. It calls mature installed media tools rather than reimplementing denoising algorithms or training another model. Jobs have explicit resource budgets, machine-readable results, source provenance, verified outputs, and resumable serial batches.
Native CI targets: Linux, Windows and macOS CPU. GPU, another MCP server, a DCC host, a network connection, and a cluster scheduler are not required by the CLI. Individual backends retain their own requirements. Render denoising currently uses an installed Blender; standalone OIDN is a planned replacement/alternative, not shipped yet.
| Operation | Backend | Initial support |
|---|---|---|
| Render denoise | Installed Blender compositor / OIDN | Opaque, flat single-part scene-linear EXR with real beauty, albedo and normal passes; float32 RGBA beauty output |
| Image denoise | Installed FFmpeg nlmeans |
Opaque 8-bit SDR image to PNG |
| Video denoise | Installed FFmpeg atadenoise |
Opaque 8-bit SDR to lossless FFV1/Matroska; copied audio; independently checked relative timestamps |
| Resize | Installed FFmpeg Lanczos | Deterministic CPU 2× interpolation; not neural super-resolution |
| Neural upscale | Optional waifu2x-ncnn-vulkan -g -1 |
Experimental adapter only; actual CPU probes failed on this cloud for official 20250915 and 20250504 releases; unavailable here |
No backend or weight is automatically downloaded. Unsupported HDR video, transparency, rotation, non-square pixels, extra subtitle/data streams, or unknown formats fail closed. Render EXR does not get tone-mapped or silently converted to PNG. Source passes and metadata remain in the retained original EXR; this first adapter produces a separate beauty master rather than claiming to preserve every channel in the derivative.
Rust 1.85+; see platform guarantees and limits. Dependencies are pinned in Cargo.lock.
CARGO_BUILD_JOBS=1 cargo build --locked
./target/debug/dcc-media-enhance doctor
./target/debug/dcc-media-enhance schema
./target/debug/dcc-media-enhance probe source.png
./target/debug/dcc-media-enhance plan examples/image.json
./target/debug/dcc-media-enhance run examples/image.json
./target/debug/dcc-media-enhance batch examples/batch.json
./target/debug/dcc-media-enhance status output.png.job.json
./target/debug/dcc-media-enhance cancel output.png.job.jsonEdit example input/output paths before running. Paths inside jobs are relative to
the job JSON file, not the caller's working directory. Output parents must already
exist. Every response is a single JSON object on stdout; backend logs are separate,
bounded local artifacts. A backend cannot be replaced by an arbitrary command in a
job. doctor observes installed capabilities and available capacity; it does not
reserve capacity or install software.
- Exit 0: success, including verified cache hit
- Exit 2: invalid or unsupported request / no-clobber conflict
- Exit 3: dependency, backend, validation, I/O, or partial batch failure
- Exit 4: resource budget exceeded
- Exit 5: cancelled
See the JSON schema, contract, backend selection and licenses, and agent skill.
The original is hashed before and after processing and never opened for writing. Workers write to private same-filesystem staging, then the output is decoded and checked. Publication uses no-clobber hard links plus fsync. The final manifest is the commit marker. Because two filenames cannot be committed in one filesystem operation, a crash can leave a complete orphan output; it is never treated as a cache hit, silently removed, or overwritten. Inspect the recorded state and use a new output path if necessary.
A cache hit requires the same effective request, input content, executable, linked runtime libraries, model files, helper/runner code and a matching output hash. OS file locks exclude concurrent writers and are released after crashes. Retrying a batch reuses verified completed items; failed items restart their enhancement stage against the preserved source. Interrupted staging is not assumed valid.
Thread hints, pixel/byte limits, disk reserve, timeout and a sampled process-group RSS watchdog are enforced. The RSS watchdog is not an OS hard memory boundary. Use an existing cgroup/container/scheduler for hard isolation. Budgets apply per subprocess; probes additionally have a 15-second/256-MiB/2-MiB result cap. Video verification is intentionally bounded and rejects an oversized probe response. Long videos will need a future streaming verification implementation.
The canonical skill lives in this repository. Existing host skills should save a raw EXR with true auxiliary passes, release the render process, then invoke this CLI as a separate job. Denoise failure must not trigger a re-render. Linear EXR is the master; display transforms and preview encodes are explicit later stages.
This is the preferred contract for future rendering workflows once the tool is
installed and verified in that environment. Existing Blender/Houdini/other host
routes have not all been migrated. The public marketplace should reference
this skill at an immutable revision after review, rather than duplicate its body.
Existing Houdini idenoise and texture/OCIO tools remain complementary adapters.
Batch JSON is portable input for an existing scheduler. The scheduler provides asset placement, machine selection, hard limits and retries; this project does not provision machines, manage a cluster, or invent another queue service.
See the measured local validation report.
Run cargo test --locked and the actual-backend
smoke tests. Decode success and synthetic regression tests are
not artistic quality benchmarks. Manifests distinguish executed checks from
not_run; enhanced derivatives are not evidence that previously unreadable detail
was factual.
The source and original adapter code are available under MIT. No versioned release or prebundled third-party runtime/model is published. Native CI builds and executes the CLI and FFmpeg regression tests on the four targets listed below; that scope does not establish render/OIDN or neural-upscale support on every platform. The historical Linux preview described in the local validation report requires glibc 2.39+ and libgcc_s; older-distro distribution needs a separately built/tested baseline artifact. A Rust executable does not remove backend ABI, native-library, model-license or supply-chain obligations. Rust dependency notices are in THIRD_PARTY_NOTICES.md. The wrapper's MIT license does not relicense FFmpeg, Blender, OIDN, ncnn, or model weights.
PR/main CI builds and runs Linux x86_64, Windows x86_64, macOS x86_64 and ARM64
binaries. Download the matching tested workflow artifact, verify its archive and
binary SHA-256, then use the executable by absolute path. No backend is bundled.
capabilities gives the machine-readable CLI surface and platform limits.
See platforms and release workflow and dcc-mcp-cli integration. Windows/macOS cache reuse is deliberately disabled until full loader identity is verified. Real FFmpeg CI tests do not establish OIDN or neural-upscale support on those machines. Existing Linux render evidence remains scoped to that host.