RESOLVE FX · TEMPORAL + SPATIAL DENOISE
Noise Reduction.
Protect motion first.
Then remove what remains.
Noise Reduction is not one strength slider. Temporal NR decides what can be averaged across time, Spatial NR cleans what remains inside one frame, and AI UltraNR adds a noise-profile workflow for severe material. The page below separates those jobs and makes the failure modes visible.
Read the denoise tradeoff+
Residual synthetic noise: 77%
Synthetic detail retention: 95%
Synthetic temporal ghost risk: 5%
Independent deterministic teaching proxy, not Resolve rendering. It visualizes the causal tradeoff between temporal evidence, motion exclusion, residual spatial cleanup, chroma treatment, ghost risk and texture loss. It does not reproduce Blackmagic motion estimation, AI UltraNR, exact thresholds, GPU precision or noise-profile math.
01THE DENOISE MODEL
Time first. Space second. Motion decides how far you can push both.
The cleanest workflow is not “remove all noise.” It is to classify what is random, what is moving, what is real texture, and which stage can remove the noise with the smallest visual cost.
Temporal first separates random noise from stable detail
Temporal NR uses neighboring frames. Where scene content remains coherent over time, random sensor noise changes from frame to frame and can be suppressed without crude spatial blur. Moving subjects cannot simply be averaged, which is why Frames Either Side, motion estimation, Motion Range, and Motion Threshold form one decision system.
Spatial is a second stage, not a fix for poor temporal tuning
Spatial NR analyzes the current frame and can work where motion limits the temporal stage. The cost is a greater risk of smoothing pores, hair, fabric, foliage, and edge texture. The practical pairing is to remove as much random noise as safely possible with Temporal NR, then use Spatial NR only for the residual pattern.
Luma and Chroma do not have to need the same strength
Split/Gang Luma Chroma exists because luminance noise and colored speckle damage the image differently. Chroma can often be cleaned more aggressively while Luma remains lower to protect fine detail. Split is not permission to eliminate chroma at any cost, because fine color texture and thin edges can disappear as well.
Blend here requires reading the label, not relying on muscle memory
The internal Temporal/Spatial Blend controls run from the fully calculated stage at 0 toward the original image at 100. Noise Reduction documents Global Blend separately: 0.0 is the Temporal + Spatial NR result and 1.0 is no noise reduction. That is unusual compared with many Resolve FX, so final dosage should be verified by A/B rather than inferred from slider position.
Key exception worth memorizing: Blackmagic documents Noise Reduction Global Blend in the opposite direction from many ordinary Resolve FX — 0.0 is the current NR result, 1.0 is no NR. Verify it by A/B rather than by habit.
02THE ACTUAL PANEL
Nine captures, five parameter groups, no mystery slider.
The native screenshots anchor every visible dropdown and section. The written model then adds only documented conditional behavior such as AI UltraNR Analyze/Noise Profile.
TEMPORAL ANALYSIS
How much time to analyze and how to interpret motion
- Frames Either Side
Chooses 0–5 neighboring frames on each side of the current frame. 0 performs no frame averaging; higher values provide more temporal samples and substantially increase processing cost.
CheckOn fast motion, start at 1. More frames make sense only when texture remains stable over time and no trail/ghost appears.
- Mo. Est. Type
Faster uses lighter motion estimation, Better excludes moving regions more accurately, and None disables motion estimation entirely.
CheckIf temporal cleanup works on the background but damages hair/hands/water, A/B Faster ↔ Better first; None is a diagnostic edge case, not a universal speed mode.
- Motion Range
Small, Medium, and Large define the motion scale the estimator should expect to exclude. This is not Noise Reduction amount.
CheckLarge helps with faster movement and broader motion blur, but it can exclude more of the image from temporal averaging.
TEMPORAL THRESHOLD
How much noise to remove and where to protect motion
- Split Luma Chroma
Separates the single Threshold into independent Luma and Chroma controls for the temporal stage.
CheckIf chroma speckle is stronger than luminance grain, keep Luma lower and raise Chroma independently.
- Threshold / Luma / Chroma
0 means no temporal noise reduction for the corresponding component; raising it increases suppression, but excessive values can remove fine detail.
CheckTune until noise stops being distracting, not until the image becomes completely sterile.
- Motion
This threshold separates pixels considered moving and excluded from temporal averaging from pixels considered static enough. Lower values protect more subtle motion; higher values let NR affect more of the frame.
CheckGhosting = lower Motion first or improve motion estimation; do not hide the problem only by reducing Threshold.
- Blend
Internal Temporal NR mix: the manual describes 0.0 as the fully calculated temporal result, moving toward 100 as returning the image without temporal NR.
CheckUse after motion/threshold tuning is correct; Blend cannot fix incorrect motion classification.
SPATIAL ANALYSIS
Algorithm and scale of single-frame analysis
- Mode
Faster, Better, and Enhanced perform spatial pattern analysis with increasing complexity. AI UltraNR adds AI-driven noise-pattern identification and exposes an Analyze/Noise Profile workflow when selected.
CheckAI UltraNR is useful on severe noise, but it should still be compared against Enhanced for texture retention rather than selected automatically because it says AI.
- Radius
Small, Medium, and Large define the neighborhood size for spatial analysis. A larger Radius can improve detail handling at high Threshold values but reduces performance.
CheckSmall + moderate threshold is a safe starting point; if strong spatial NR makes detail chunky/aliased, try Medium/Large and re-check texture.
SPATIAL THRESHOLD
Residual cleanup without plastic faces
- Split Luma Chroma
Provides independent spatial Luma/Chroma thresholds when colored noise needs stronger treatment than luminance texture.
CheckWatch pores, eyelashes, hair, and fabric: Luma is usually where excessive spatial smoothness becomes visible first.
- Threshold / Luma / Chroma
Sets the strength of spatial reduction inside the current frame. 0 means no suppression; increasing it removes more high-frequency noise while raising the risk of texture loss.
CheckIf Spatial Threshold must be pushed hard, revisit the temporal stage; often more noise can be removed there more safely.
- Blend
Internal wet/dry for the spatial stage: 0.0 shows the calculated spatial result, while increasing toward 100 restores the original image.
CheckThis is local dosage for Spatial NR, not the final mix of the whole FX.
AI ULTRANR + GLOBAL
Noise profile and final dosage
- AI UltraNR · Analyze / Noise Profile
AI UltraNR locates a very flat image region for noise-pattern estimation, displays a sample patch, and derives a Luma/Chroma noise profile. The patch can be moved/resized, but the automatic selection is the normal starting point.
CheckDo not place the patch on a face, hair, fabric texture, or an edge just because it looks dark; the sample should contain minimal scene detail.
- Global Blend
For Noise Reduction, Blackmagic documents the scale separately: 0.0 is the image with the current Temporal + Spatial NR settings, while 1.0 is the image with no noise reduction. This is opposite to the familiar behavior of many other Resolve FX.
CheckVerify with bypass/A-B. If the slider is needed to hide ghosting or waxy texture, fix the responsible temporal/spatial stage first.
Do not copy a number from another camera or shot. Thresholds are content-dependent; Motion parameters are movement-dependent; AI UltraNR noise profile is sample-dependent. The transferable skill is the order of diagnosis, not a preset.
03PROTECT DETAIL BEFORE CLEANUP
Temporal classification first, spatial cleanup second.
This sequence mirrors Blackmagic’s own practical guidance: establish Temporal NR, protect motion, then bring in Spatial NR only for what remains.
Read the noise before touching NR
At 100%, identify whether the problem is mostly random luma grain, chroma speckle, compression noise, or a different artifact such as banding.
Teaching workflow, not a literal Resolve node graph. The important order is diagnose → protect motion → set temporal strength → clean residual spatial noise → verify the final mix and codec.
Related tools: Frame Replacer ↗ · Deband ↗ · Object Removal ↗
04BUILD IT IN RESOLVE
Four exercises that teach the real tradeoffs.
These are diagnostic training procedures around documented controls, not universal settings. Use footage with visible texture and judge it at 100% in motion.
High-ISO interview with a moving face
Remove random noise from the static background without turning skin/hair into smear.
PrepareUse a clip with a face, micro-expression, and a calm background; judge the Viewer at 100%.
- 01
Set Frames Either Side = 1; compare Faster and Better.
- 02
Choose Motion Range from actual head/hand speed, not from noise level.
- 03
Raise Temporal Threshold until the background cleans up, then adjust Motion if hair begins to trail.
- 04
Add minimal Spatial NR only for residual speckle.
The background is visibly cleaner while pores, eyelashes, and hair contours retain motion and texture in playback.
Chroma noise stronger than Luma
Remove purple/green speckle without over-smoothing luminance detail.
PrepareUse underexposed material where colored speckles are more obvious than luminance grain.
- 01
Establish baseline Temporal NR with linked Threshold.
- 02
Enable Split Luma Chroma and keep Luma at the minimum useful level.
- 03
Raise Chroma separately until distracting color speckle recedes.
- 04
Apply the same logic in Spatial Threshold only if needed.
Colored noise is reduced more aggressively than luminance noise while micro-contrast and skin texture avoid a smeared look.
Noisy fast motion
Avoid ghosting while trying to remove noise from a moving subject.
PrepareChoose a handheld/action shot with fast movement, motion blur, and a textured background.
- 01
Start with 1 frame and Better; test Motion Range in Medium/Large.
- 02
Keep Motion Threshold low enough that subtle moving pixels are excluded from averaging.
- 03
Do not add Frames just to make one still cleaner; judge playback.
- 04
Use moderate Spatial NR for noise the temporal stage intentionally leaves on moving regions.
Motion keeps a normal trajectory without echo/trail while residual noise remains natural rather than pumping.
Extreme underexposure and AI UltraNR
Determine whether AI UltraNR actually beats Enhanced on severe sensor noise.
PrepareUse a severe high-ISO/underexposed shot with one reasonably flat area and important fine detail.
- 01
First get acceptable Temporal NR; AI UltraNR does not replace the motion workflow.
- 02
Compare Spatial Mode Enhanced and AI UltraNR on the same frame.
- 03
In AI UltraNR, run Analyze, inspect the sample patch and noise profile; do not move it onto a textured object.
- 04
Compare eyes/hair/fabric and playback, not only an empty shadow area.
Choose the mode that produces less noise with less loss of meaningful texture; AI is not considered the winner automatically.
05CHOOSE THE FAILURE MODEL
Noise, banding, corruption and object removal are different repairs.
Do not make Noise Reduction solve defects that are not random noise. Choosing the right failure model preserves more detail than simply pushing thresholds harder.
| FX | What it changes | Use it when |
|---|---|---|
| Frame Replacer ↗ | Whole-frame reconstruction of one damaged frame. | When a frame is physically corrupt/flash/glitched rather than merely noisy. |
| Noise Reduction | Temporal + spatial suppression of random sensor/compression noise. | When the detail exists but is contaminated by a random noise pattern. |
| Object Removal ↗ | Reconstruction of a masked region from background context. | When a specific object must be removed rather than noise across the whole frame. |
| Deband ↗ | Repair of contour bands/steps in smooth gradients. | When the problem is posterization/banding rather than random pixel noise. |
| Detail Recovery ↗ | Recovers/enhances high-frequency detail from another branch. | When useful detail needs recovery after processing, but not as a substitute for correct NR. |
06DIAGNOSE THE ARTIFACT
Ghost, wax, pumping and wrong mix have different causes.
Noise reduction becomes predictable when you diagnose the stage that failed. Do not solve temporal artifacts with spatial smoothing or texture loss with a global mix before fixing the source stage.
01A moving subject leaves a ghost/trail+
Likely cause: Temporal NR includes too many moving pixels: Motion Threshold is high, Motion Range/Estimation mismatches the scene, or too many Frames are used.
Fix: Reduce Frames to 1, A/B Better, match Motion Range, and lower Motion Threshold until the trail disappears; only then restore Threshold.
02Skin looks waxy and hair clumps together+
Likely cause: Spatial Luma/Threshold is too high or Radius/Mode is over-smoothing; the temporal stage may also be underused.
Fix: Reduce Spatial Luma, revisit Temporal NR and move safe cleanup there; Chroma can remain higher than Luma for colored noise.
03Noise disappears and returns during motion+
Likely cause: The temporal motion mask correctly excludes moving regions, but Spatial NR does not clean enough residual noise or Motion threshold changes coverage too abruptly.
Fix: Stabilize motion classification and add moderate Spatial NR for residual noise instead of forcing Temporal NR to clean the entire moving subject.
04AI UltraNR analyzes the “wrong” patch+
Likely cause: Automatic flat-area search selected an unexpected region, or the patch was manually moved onto textured content.
Fix: Evaluate the auto result first. If moving the patch, choose a genuinely flat noise sample without edges/texture; then re-check Luma/Chroma profile and detail retention.
05Global Blend feels “backwards”+
Likely cause: For Noise Reduction, the manual documents 0.0 as processed NR and 1.0 as no NR — opposite to the habit learned from many other FX.
Fix: Perform an explicit A/B: temporarily make NR obvious, verify both slider extremes, then use Global Blend only as final trim.
06Noise looks different after export than in the Viewer+
Likely cause: Viewer scale/cache/proxy and delivery codec change the appearance of fine noise; compression can turn residual grain into macroblocking/mush.
Fix: Approve at 100%, verify full-res/cache state, and make a short render in the real delivery codec before final approval.
07BEFORE DELIVERY
Approve motion, texture and the encoded noise floor.
A still frame can look beautifully clean while playback contains trails and the delivery codec turns the remaining fine noise into mush. Final QC must include both.
08FAQ
Practical Noise Reduction questions.
Why usually start with Temporal NR?
Because it uses temporal redundancy and can remove random noise in stable regions with less spatial-detail loss. Spatial NR then cleans what the temporal stage leaves because of motion.
How many Frames Either Side should I use?
There is no universal maximum. On fast motion, 1 is often safer; more frames help on stable footage but cost more GPU/time and can increase temporal artifacts.
How is Better different from Faster?
Better performs more accurate motion estimation and can exclude moving regions more precisely, at higher cost. Check hair, hands, water, foliage, and motion blur.
When should Split Luma Chroma be enabled?
When chroma speckle and luminance noise differ meaningfully. Chroma can often be raised above Luma while preserving fine luminance detail.
Is AI UltraNR always better than Enhanced?
No. It targets real-world noise patterns and is particularly useful on severe noise, but the final choice depends on detail retention, motion, and render cost on the actual footage.
Why does Noise Reduction Global Blend look reversed?
Blackmagic documents this specific Global Blend as 0.0 = current Temporal + Spatial NR and 1.0 = no noise reduction. Do not automatically transfer the habit learned from other Resolve FX.
Sources and boundaries
The page separates native interface evidence from teaching models. Visible control names and dropdown states come from nine Resolve 21 captures in the project archive. Temporal/Spatial semantics and the Noise Reduction-specific Global Blend direction are cross-checked against Blackmagic Resolve FX documentation; AI UltraNR behavior is cross-checked against Blackmagic’s Resolve 19 New Features and contemporary Colorist training material. The animated lab is an independent deterministic proxy and does not claim renderer equivalence.
Verify Studio availability, GPU performance and any UI differences in the Resolve build you actually use. The practice blocks are explicit training procedures around documented controls and captured states; they were not certified by a live Resolve render during this site-only update.