OpenNR Denoise is a free, open-source (MIT) OpenFX plugin that brings serious spatio-temporal video noise reduction to the free version of DaVinci Resolve — the tool free-version users are missing most, since Resolve's built-in Noise Reduction palette is Studio-only.
It runs on the Color page (OpenFX panel), the Edit page (Effects Library), and the Fusion page, is fully GPU-accelerated (Metal / CUDA / OpenCL), and is built around something even the Studio NR tool doesn't offer: transparency. It measures your footage's noise automatically and shows you everything — the measured levels, what's being removed, and where each stage is working.
| Noisy input | OpenNR output (default settings) |
|---|---|
 |
 |
Validation scene, 5% Gaussian noise: 26.0 dB → 39.0 dB PSNR at defaults. Hard edges, fine checker texture and 2-px lines survive.
| Noise Analysis view (live measurements) | Temporal Activity view (green = averaging) |
|---|---|
 |
 |
- Auto Setup — one button sets up the whole plugin (v3). It analyzes several frames spread across your clip (noise level, chroma character, spatial correlation, camera motion) and writes the best settings into the visible sliders — you immediately see what it chose and dial anything in or back from there. Not a mode: afterwards everything is ordinary manual state. One Cmd+Z (or the Revert button) restores everything, a read-only Analysis line reports what it measured, and your Step 4 look choices are never touched.
- Self-tuning Auto Setup (v3.5) — the class table is only the starting point: the button now test-denoises a centre crop of your actual clip at a 3×3 grid of (temporal, spatial) strengths and keeps the statistically best pair, scored by Stein's unbiased risk estimate — a measurement of true denoising error that needs no clean reference. Deterministic, ~3–5 s at button press, and the Analysis line says "· tuned" when it ran.
- Automatic noise profiling, done right for video — three estimators run per frame: frame-to-frame differences (immune to the spatially-correlated noise that compression and debayering create), plus fine- and coarse-scale spatial estimators. Robust medians/quantiles keep motion, edges and texture out of the measurement. Lock Profile (fixed in v3.4) freezes exactly the measurement you are looking at — the playhead frame, confirmed by a tracked sweep of its neighbours — so every frame filters against the numbers you approved. The box turns blue with a padlock while locked, the HUD shows LOCKED, and it survives save/reload.
- Auto Region (v3.4) — one click finds the flattest patch on the current frame (scored against the expected noise at its brightness, so dark noisy flats aren't penalized), moves the sampling box there and shows it. Watch the live measurement, then lock it.
- Brightness-aware temporal gate (v3.4) — the measured noise-vs-brightness curve now calibrates the temporal knee, Ghost Guard and the motion-search thresholds per pixel, like the spatial stage always did: shadows (noisier) keep their averaging instead of reading as motion, highlights gate tighter. Measured +1.35 dB in the shadows on a brightness-profiled scene.
- Motion-tracked temporal NR (v3, ~±8 px since v3.3) — before gating each neighbour frame the temporal stage re-aims its patch with a hierarchical search (step-4 coarse grid on block means, then a converging ±1 refine walk), so real pans and handheld drift keep their across-frames averaging. The hard-knee gate is unchanged — still exactly zero past the knee, still ghost-proof — and shifted matches face a steeper gate.
- Render Boost (v3.5) — during sequential playback and exports, the previous frame's temporal result joins the merge as one extra gated candidate, so static areas compound frame over frame — up to about twice the effective frames. The knee actually gets SHARPER against the cleaner history (a diff against an n-frame average has less spread than a frame-pair diff), so the anti-ghosting guarantees hold: a panning test chain is bit-identical to boost-off. Scrubbing, parameter changes and clip edges fall back automatically. Off by default; +1.94 dB on a 6-frame static chain for ~1 ms at HD.
- Exposure match (v3.5) — auto-exposure steps, iris pulls and light flicker used to read as motion and knock whole frames out of the stack. Each neighbour now gets a measured global exposure offset before the motion tests, so a matched patch under flicker blends like a static one while the output keeps your frame's exposure. Flicker-free footage is bit-exact with v3.4.
- Firefly removal (v3) — single-frame impulses (hot pixels, sensor fireflies) are clipped to the 3-frame temporal median. Three independent tests must agree before a pixel is touched, so thin fast-moving detail survives.
- Detail Rescue (v3.1) — smoothing without blur. After the spatial stage, anything it changed by more than a noise-sized amount is put back. Crank the strengths as hard as you like: flats and faces go buttery while edges physically cannot blur. Auto Setup raises it with the noise class.
- Noise EQ (v3, labeled in v3.1) — the spatial stage split into bands by the SIZE of the noise: Fine Grain (~1 px), Clumps (3–8 px), Stains (16 px+) and Color Blotches, grouped under a "Cut Noise By Size" section. Touch any band and the EQ scope pops up in the viewer — one lane per band, the bar is how much you're cutting, the amber line is how much noise was measured at that size.
- Extended ranges (v3.1) — Strength to 3, spatial strengths to 150, EQ Fine to 300 (its top half is no longer a silent no-op), search radius to 10, band sliders to 150 (extra drive widens tolerance and reach).
- Deband + noise matte (v3) — gradient-aware debanding that never touches real edges, and a Matte: Noisiness view that writes the noise-dominance map to RGB + alpha for keying downstream.
- Per-step scopes (v3.1) — checkboxes where you need them: Step 1 Measurements (noise levels, residual, SNR gain, brightness curve, histogram — rebuilt with 2× text so it stays readable at fit zoom), Step 2 Motion Map (live mini-map: green = frames stacking, red = protected), Step 3 Noise EQ. Scopes draw over any view and combine freely. Noise Removed shows exactly what is being taken out, auto-gained to the measured noise level with a soft knee. Resolve's own NR is a black box — this one shows its work.
- Step-by-step controls — numbered sections (1 · Measure, 2 · Temporal, 3 · Spatial, 4 · Refine, 5 · Inspect) with per-stage Enable toggles, and short plain-English tooltips on every control.
- Temporal NR — motion-adaptive averaging over 3 or 5 frames, gated against the measured noise floor. No ghosting: moving pixels fall back to spatial filtering automatically.
- Spatial NR — noise-adaptive non-local means (or faster bilateral), separate luma / chroma strengths, edge-aware Preserve Detail.
- Profile from a region — restrict measurement to a rectangle you place on a flat area (or let Auto Region place it). The box is a live probe: yellow and draggable while measuring per frame, dim blue with a padlock once locked. Or go full manual.
- Works in DaVinci Resolve free and Studio (17+), macOS binaries are universal (Apple Silicon + Intel, macOS 11+).
Measured on an Apple M1 Max (test/bench_metal.mm), v3.5, best of repeated
runs. The bench feeds realistic frames (scene + per-frame noise). The
non-boost paths are byte-identical math to v3.4 (parity-verified) — deltas
vs the old table are bench-session variance, not code. Panning costs more
than v3.2 because the motion search genuinely re-aims up to ~8 px; toggle
Motion Tracking off for the old speed. Lock Profile is also the fast path:
with no scopes open, a locked profile skips the whole input measurement.
Render Boost is measured over sequential frames (the export case), history
valid every frame.
| Resolution | Better (NLM, R3, 5 frames) | Locked profile | Panning footage | 7 Frames | Render Boost (sequential) | Faster (bilateral, R2, 3 frames) |
|---|---|---|---|---|---|---|
| HD 1920×1080 | 10.2 ms (98 fps) | 8.5 ms (118 fps) | 21.9 ms (46 fps) | 11.4 ms (88 fps) | 11.0 ms (91 fps) | 6.2 ms (162 fps) |
| UHD 3840×2160 | 38.7 ms (26 fps) | 33.3 ms (30 fps) | 84.2 ms (12 fps) | 42.5 ms (24 fps) | 41.2 ms (24 fps) | 22.4 ms (45 fps) |
Real-time UHD at maximum quality on Apple Silicon; Deep Clean adds ~3 ms at HD when enabled, Render Boost ~0.7 ms over the same-settings static row.
- Download
OpenNR-3.7.0-macOS.pkgand double-click it. It's signed and notarized by Apple, so it just opens — no right-click, no "Open Anyway", no Terminal. - Restart DaVinci Resolve. If you had a previous OpenNR version, also delete
Resolve's plugin scan cache so it rescans:
rm ~/Library/Application\ Support/Blackmagic\ Design/DaVinci\ Resolve/OFXPluginCacheV2.xml
Alternative (no installer): unzip OpenNR-3.7.0-macOS.zip and double-click
Install OpenNR (macOS).command, or copy OpenNR.ofx.bundle into
/Library/OFX/Plugins/ yourself. The bundle carries a stapled notarization
ticket, so a copied or AirDropped one loads without clearing any quarantine
flag.
Download OpenNR-x.y.z-Windows.zip from the latest release, unzip, and either
double-click Install OpenNR (Windows).bat or drag OpenNR.ofx.bundle into
C:\Program Files\Common Files\OFX\Plugins yourself. Restart Resolve.
(The Windows build renders through OpenCL — works on NVIDIA, AMD and Intel
GPUs. A CUDA-native path exists in the tree, gated until verified on hardware.)
Copy OpenNR.ofx.bundle into /usr/OFX/Plugins/.
- Color page → Effects → OpenFX → Filters → OpenNR → OpenNR Denoise — drag onto a node
- Edit page → Effects Library → OpenFX → Filters → OpenNR — drag onto a clip
- Fusion page → Effects Library → OpenFX → OpenNR — add as a tool
- Drop Hush Open NR on a node (Color page) or clip (Edit page) and click Auto Setup — it measures the clip and dials in every slider. Adjust Strength to taste; one undo reverts the whole setup.
- Check the measurement: tick Step 1 → Scope: Measurements. If the numbers read implausibly low for obviously noisy footage, raise Auto Profile Adjust — or measure from a flat area with Automatic (From Region).
- Check what's being removed: Step 5 → View → Noise Removed. Featureless static = good. Visible edges/faces = too aggressive: raise Preserve Detail or Detail Rescue (which lets you keep the smoothing AND the edges).
- Untick the scopes, set View back to Result. Done.
Colorists: apply NR early in the node tree (pre-grade), as with Resolve's built-in NR — denoise before contrast/saturation expansion amplifies the noise.
A/B each stage: the Enable checkboxes in Steps 2 and 3 toggle temporal and spatial filtering independently — flip them while watching the viewer (or the Temporal Activity view) to understand what each is contributing.
| Step | Control | What it does |
|---|---|---|
| — | Strength | Overall amount; scales every strength at once. 0 = off, 1 = normal, up to 3 the filters widen what they treat as noise. |
| — | Auto Setup (Analyze Footage) | Analyzes the clip (region-aware: sigmas come from your tracked region at the playhead, motion from frames across the clip) and writes the best settings into the sliders below, then locks the profile. Everything stays manually adjustable; one undo reverts. |
| — | Auto Region (Find Flat Patch) | Scans the current frame for the flattest area, switches the profile to From Region and moves the yellow box there. Watch it live, then Lock Profile (or run Auto Setup). One undo restores. |
| — | Analysis | Read-only report of what the last Auto Setup / Auto Region / Lock measured and decided. |
| — | Clean Slate (All Off) | Zeroes every processing control — true passthrough, build up manually from nothing. One undo restores. |
| 1 | Noise Profile | Automatic (whole frame) / Automatic (from region) / Manual. Unlocked, the measurement is live per frame; Lock Profile is the "stop changing" switch. |
| 1 | Region Center X/Y, Size | The measurement rectangle (drag it in the viewer, or click Auto Region). Put it on a flat area. Yellow = sampling live every frame; blue + padlock = frozen. |
| 1 | Auto Profile Adjust | Scales the measurement (×0.25–×6) — live and locked profiles alike. |
| 1 | Manual Luma / Chroma Noise (%) | Direct noise levels, used only in Manual mode. Clean ≈ 0.5–1, noisy ≈ 2–5, low-light ≈ 5–10. |
| 1 | Lock Profile | Freezes the measurement at the playhead — exactly what you are looking at, confirmed across a few tracked neighbouring frames (a subject crossing the patch is rejected). Park on a frame where it looks right, then lock. Saved with the project. |
| 1 | Scope: Measurements | The measurement panel, drawn live in the viewer. |
| 2 | Enable Temporal NR | Toggle the across-frames stage. |
| 2 | Number of Frames | 3, 5 or 7 frames averaged. |
| 2 | Motion Tracking | Hierarchical shift-search patch matching (~±8 px) so real pans keep their temporal averaging. |
| 2 | Luma / Chroma Strength | Temporal blending (0–125; above 100 matching neighbours outweigh the current frame). |
| 2 | Motion Threshold | How much change counts as motion (0–150). Seeing smear on movement? Lower this first. |
| 2 | Ghost Guard | Second motion test on the signed patch mean — catches the slow, subtle motion (slow-mo smear) the magnitude gate can't see. Nearly free; leave on. |
| 2 | Firefly Removal | Clips single-frame impulses (hot pixels) to the temporal median. |
| 2 | Scope: Motion Map | Live mini-map: green = frames stacking, red = motion-protected. |
| 3 | Enable Spatial NR | Toggle the within-frame stage. |
| 3 | Method | Better (NLM, patch-based) or Faster (bilateral). |
| 3 | Search Radius | How far to look for similar patches (1–10 px). |
| 3 | Luma / Chroma Strength | Spatial filtering (0–150; above 100 the filter widens what counts as noise). |
| 3 | Preserve Detail | Edge-aware protection of real structure. |
| 3 | Detail Rescue | Restores anything the smoothing changed by more than a noise-sized amount — crank strengths, keep edges. |
| 3 | Noise EQ · Cut Noise By Size | Fine Grain ~1 px (to 300) / Clumps 3–8 px / Stains 16 px+ / Color Blotches (each to 150). Touching one pops the EQ scope. |
| 3 | Scope: Noise EQ | Per-band panel: bar = your cut, amber line = measured noise at that size. |
| 4 | Shadow Desaturate, Desat Range | Saturation-vs-luma curve to hide chroma noise in shadows. |
| 4 | Luma Texture | Re-injects original brightness texture after denoising. |
| 4 | Deband | Gradient-aware banding smoother + micro-dither. Edges are never touched. |
| 4 | Film Grain, Grain Size, Grain Color | Clean synthesized grain to finish. |
| — | Global Blend | Final crossfade original ↔ processed result, after everything else. |
| 5 | View | Result / Split / Input / After Temporal / Noise Removed (auto gain) / Noise Analysis / Temporal Activity / SNR Map / Matte: Noisiness (map in RGB + alpha, for keying downstream). |
| Studio NR control | OpenNR equivalent |
|---|---|
| Temporal NR → Frames | Step 2 → Number of Frames |
| Temporal NR → Motion Est. Type | (automatic; difference-based gating) |
| Temporal NR → Motion Range / Threshold | Step 2 → Motion Threshold |
| Temporal Threshold → Luma / Chroma | Step 2 → Luma / Chroma Strength |
| Spatial NR → Mode (Faster/Better) | Step 3 → Method |
| Spatial NR → Radius | Step 3 → Search Radius |
| Spatial Threshold → Luma / Chroma | Step 3 → Luma / Chroma Strength |
| Blend | Strength (inverted sense) |
| — | Step 1 (noise measurement) and Step 4 (analysis views) — no Studio equivalent |
- No motion-compensated temporal NR. Studio's temporal NR tracks motion vectors and averages along motion paths; OpenNR detects motion and protects it instead (spatial NR takes over there). Comparable on locked-off/slow shots; Studio keeps an edge on fast pans.
- No AI model. Studio 20+ adds UltraNR; OpenNR is classical (NLM) — very competitive at normal noise levels, less magical on extreme starlight footage.
- Requires float RGBA processing (all Resolve versions provide this).
macOS (Xcode command line tools):
cd plugin && make # universal arm64 + x86_64, macOS 11+
sudo make install # copies to /Library/OFX/PluginsLinux (CUDA toolkit + OpenCL headers):
cd plugin && make && sudo cp -r OpenNR.ofx.bundle /usr/OFX/Plugins/Windows: build the four sources (OpenNRPlugin.cpp, CudaKernel.cu,
OpenCLKernel.cpp, OFX Support lib) into OpenNR.ofx with MSVC + nvcc,
mirroring Blackmagic's OpenFX samples. Prebuilt Windows/Linux binaries via CI
are on the roadmap.
Release packaging (macOS): ./build_release.sh produces the .pkg and .zip
(sets the deployment target, ad-hoc signs, and sanity-checks the minimum-OS
stamp).
The algorithm is validated outside Resolve; run these before any release:
test/test_denoise.cpp— synthetic scenes with known injected noise: estimator accuracy on iid and spatially-correlated noise (the real-world camera/compression case), motion robustness, PSNR gates (+13 dB static / +9 dB motion / +11 dB correlated at σ=0.05), identity at Strength 0 and with both stages disabled, duplicate-frame guard, region profiling, and renders of every view mode (including the HUD) for visual inspection.test/test_metal.mm— GPU parity: the real Metal pipeline vs. the CPU reference across 12 configurations (agreement ~2×10⁻⁵).test/bench_metal.mm— the performance numbers above.
cd test
c++ -O2 -std=c++14 -I../plugin test_denoise.cpp -o test_denoise && ./test_denoise
c++ -O2 -std=c++14 -I../plugin test_metal.mm ../plugin/MetalKernel.mm \
-framework Metal -framework Foundation -o test_metal && ./test_metal- Measure — three per-frame estimators: median |frame difference| (measures the true total noise, immune to the spatial correlation that debayer/compression create; motion-robust via a low-quantile cross-check and a spatial-estimate clamp), median |Laplacian| at fine scale, and at coarse scale (catches blotchy correlated noise). The temporal stage is gated by the temporal estimate; the spatial stage by the spatial one.
- Temporal — each pixel is compared patch-wise (3×3) against up to ±3 frames; differences are bias-corrected by the expected noise difference and gated softly in units of σ — since v3.4 the local σ, scaled by the measured noise-vs-brightness curve, so shadows gate as generously as their real noise deserves. The achieved effective sample count is carried forward per pixel (and visualized in Temporal Activity).
- Spatial — non-local means on the temporal result, filter strength tied to the remaining noise (σ/√effN), patch distances bias-corrected by 2σ², luma-guided chroma weights, edge-aware strength reduction.
- Inspect — the analysis views are rendered by the same GPU kernels, straight from the measurement buffers (histograms included), so what you see is exactly what the filters use.
MIT — free for any use, including commercial. See LICENSE. The vendored OpenFX headers/support library are BSD-licensed by the Open Effects Association.