Moonlight AI Settings Overview
How It Works
MoonLightAI processes your game screen in real-time through a multi-stage pipeline. Understanding this pipeline helps you know which settings affect what and in which order.
Capture → Preprocessing → Detection → Filtering → Tracking/Prediction → Movement Calculation → Humanization → Device OutputCapture
A frame is grabbed from your capture card, NDI stream, or screen
Settings → Capture
Preprocessing
The frame is cropped, color-converted, and optionally GPU-accelerated
Settings → Capture
Detection
The AI model runs inference to find targets (bounding boxes or keypoints)
Settings → Detection/Model
Filtering
False positives are removed using color, shape, or template matching rules
Settings → Filtering
Tracking
Kalman filters predict target movement and compensate for latency
Settings → Tracking
Movement Calculation
Speed, smoothing, and distance scaling are applied to produce a raw delta
Aim, Flick & Trigger, RCS
Humanization
The raw delta is modified to look like natural human mouse movement
Humanizer
Device Output
The final movement is sent to your mouse via software or hardware device
Settings → Mouse/HID
Key concept — Weapon Classes: A single config file can contain multiple profiles called Weapon Classes (e.g., "Rifles," "Snipers," "Shotguns"). Settings on the Aim, Flick, RCS, and Humanizer pages are stored per-class. Settings on the Settings page are global. When you switch weapon class — manually, by hotkey, or automatically via OCR/Template Matching — all per-class settings update instantly.
Quick Start Guide
If this is your first time launching MoonLightAI, the First-Run Wizard will walk you through the essentials. Here is a summary of the steps:
1. Set Up Capture
Go to Settings → Capture and select your Capture Mode:
Capture Card — Select your device, codec, format, and resolution.
NDI — Ensure the NDI source is broadcasting on your local network.
Screenshot — No configuration needed; suitable for testing only.
Enable Crop for significantly better performance. The system only needs to process the center of your screen.
2. Load a Model
Go to Settings → Detection/Model:
Select a Model File from the dropdown (
.onnxor.encfiles in themodels/folder).Set ONNX Provider to
gpu(DirectML) for NVIDIA/AMD, oropenvinofor Intel.Set Confidence to around 50–60% to start.
Choose the correct Model Type (Object Detection or Pose Estimation) to match your model.
3. Verify Detection
Set Select Visual Type to "Debug - Crop Only" on the Capture tab.
Press Start in the sidebar.
You should see bounding boxes drawn around detected targets in the debug window.
If nothing appears, lower Confidence or check your Capture settings.
4. Configure Aim
Go to the Aim page:
Enable Aim Enable.
Set your Aim Key (e.g., Right Click).
Choose an Aim Type (Aim V2 is recommended for new users).
Set Smooth X / Y to around 25–35 as a starting point.
Set Aim FOV.
5. Test and Tune
Enter a game, hold your aim key, and observe the behavior. From here:
If aim feels too fast/snappy → increase Smooth X/Y values.
If aim feels too slow/floaty → decrease Smooth X/Y values.
If aim is jittery → increase Tracking Deadzone.
Aim Page
The Aim page is the primary interface for configuring assistive pointing behavior. It is divided into three sections: Global Settings, Targeting & Position, and a dynamic Aim Controls panel that changes based on the selected algorithm.
Global Aim Settings
These settings apply universally regardless of the selected aim algorithm.
Activation Controls
Aim Key
Multi-select dropdown
Bind one or more keys (Left Click, Right Click, Side Buttons) to activate the aim function. Multiple keys can be active simultaneously.
Aim Enable
Checkbox
Master toggle for the entire aim system. All other settings are ignored when this is off.
Timing & Limitations
Reaction Time
Slider (ms)
Artificial delay between key press and the start of aim movement. Simulates human reaction speed. Higher values are safer but feel less responsive.
Max Aim Time
Checkbox + Slider (s)
When enabled, sets a hard time limit for how long aim assist stays active per key press. Prevents the system from tracking a target for suspicious durations.
Recommended starting values: Reaction Time 80–150 ms. Max Aim Time enabled, 2–4 seconds.
Targeting & Position
Defines the search area and static offsets applied to the aim point.
Field of View (FOV)
Aim FOV
Slider
Radius of the circular search area around your crosshair, in pixels. Only targets inside this circle are tracked. Larger values find targets further from center but feel less natural.
FOV Scale
Slider
Multiplier that adjusts the FOV radius relative to your screen resolution. Use this to keep FOV behavior consistent across different monitor sizes.
Dynamic FOV
Enable
Checkbox
Toggles dynamic FOV resizing. When enabled, the search area shrinks and expands based on tracking conditions or target distance.
Min FOV
Slider
Lower bound of the dynamic FOV range. The search circle will never shrink below this value.
Max FOV
Slider
Upper bound of the dynamic FOV range. The search circle will never grow above this value.
Offsets
X Offset
Slider (px)
Fixed horizontal pixel adjustment to the aim point. Positive values shift right; negative values shift left.
Y Offset
Slider (px)
Fixed vertical pixel adjustment. Commonly used for manual recoil compensation or aiming at a specific body height (e.g., head vs. chest).
Aim Controls
This panel changes its available settings based on the Aim Type selected in the header.
Header Selection
Aim Type
Dropdown
Selects the movement algorithm: Legacy, Aim V2, or GAN. Each mode exposes different sliders below.
Aim Mode
Dropdown
Determines key behavior — Hold (active while key is held) or Toggle (press to activate, press again to deactivate).
Legacy Mode
A simple linear speed/acceleration algorithm. Best for users who want direct, predictable control.
X Speed
Slider
Base horizontal tracking speed.
Y Speed
Slider
Base vertical tracking speed.
Speed Modifier 1
Slider
Secondary speed adjustment for fine-tuning acceleration curves.
Speed Modifier 2
Slider
Tertiary speed adjustment for specific zone behaviors.
Aim V2 Mode
Advanced smoothing with distance-aware scaling. Recommended for most users.
Smooth X
Slider
Horizontal smoothing factor. Higher values = slower, smoother tracking. Lower values = snappier movement.
Smooth Y
Slider
Vertical smoothing factor. Independent from X to allow different horizontal/vertical feel.
Distance Scale Factor
Slider
Adjusts how smoothing changes with target distance. Lower values make aim faster up close and slower at range. Higher values keep speed consistent regardless of distance.
Smooth Modifiers
List editor
Define custom smoothing zones — specific smoothing values that trigger only within certain FOV ranges. For example: stronger smoothing when the target is within 50px, lighter smoothing beyond 200px.
Note: Distance Scale Factor only functions when Legacy Distance Smoothing is enabled in Settings → Mouse/HID. See Sensitivity Normalization for details.
GAN Mode
Uses a trained neural network to generate human-like movement trajectories instead of mathematical curves.
GAN Speed
Slider
Overall speed of the neural network's path generation.
GAN Humanization
Slider
Amount of noise/imperfection added to the path to mimic human error.
GAN Smoothness
Slider
Curvature of the generated trajectory. Higher values produce rounder, more flowing paths.
GAN Overshoot %
Slider (%)
Probability that the aim intentionally overshoots the target before correcting back. Adds realism to flick-like motions.
GAN Model
Dropdown
Selects which trained model file (.pth or .onnx) to use. See ML Mouse Model Tab for training your own.
Flick & Trigger Page
This page combines configuration for two distinct behaviors: Flick (rapid target acquisition) and Trigger (automated firing). The page is divided into Flick Settings and Trigger Settings.
Flick Settings
Flick is designed for faster, snap-like movements compared to the smoother tracking of the Aim page. Use it for quick target switches or one-shot weapons.
Targeting & Position
Flick FOV
Slider
Radius of the search circle. Typically set larger than Aim FOV to acquire targets further from the crosshair.
X Offset
Slider (px)
Horizontal pixel adjustment for the final flick point.
Y Offset
Slider (px)
Vertical pixel adjustment. Often used to bias toward headshots.
Flick Controls
Activation:
Flick Enable
Checkbox
Master toggle for the flick system.
Flick Key
Dropdown
Selects the activation key (Right Click, Side Buttons, etc.).
Motion Settings:
Reaction Time
Slider (ms)
Delay before the flick movement begins.
Distance Scale Factor
Slider
Adjusts smoothing based on target distance. Same behavior as Aim V2's Distance Scale Factor.
Smooth X
Slider
Horizontal smoothing. Lower = faster, snappier flicks. Higher = more controlled.
Smooth Y
Slider
Vertical smoothing. Independent from X.
Smooth Modifiers
List editor
Custom smoothing zones based on FOV ranges. Same concept as Aim V2's Smooth Modifiers.
Trigger Settings
The Triggerbot automatically fires when a target enters the crosshair or a defined region.
Trigger Controls
Activation:
Trigger Enable
Checkbox
Master toggle for auto-fire.
Trigger Key
Dropdown
Key that must be held for the triggerbot to be active.
Detection & Timing:
Trigger Scale
Slider
Adjusts the effective detection zone tightness. Lower values require the crosshair to be more precisely on-target before firing.
1st Shot Delay
Slider (ms)
Delay before the first shot after a target is detected. Simulates human reaction time.
Multi-Shot Delay
Slider (ms)
Delay between subsequent shots. Use for semi-automatic weapons or tap-fire control.
Recommended starting values: Trigger Scale 0.3–0.5. 1st Shot Delay 100–200 ms. Multi-Shot Delay 50–100 ms.
Recoil Control System (RCS) Page
The RCS page manages weapon recoil mitigation. Unlike other pages, it provides direct access to control logic and specialized modes for handling recoil patterns. The page is divided into four sections: RCS Controls, Mode Settings, Pattern Settings, and Pattern Visualization.
RCS Controls
Core activation and timing for the generic pull-down system.
RCS Enable
Checkbox
Master switch for recoil control.
RCS Strength
Slider
Vertical pull-down force to counteract muzzle rise. Higher values = stronger downward compensation for high-recoil weapons.
RCS Activation Delay
Slider (ms)
Pause after firing before recoil control engages. Useful for weapons with a stable first shot before the recoil pattern begins.
Max Time
Slider (ms)
Maximum duration for RCS during a continuous spray. Prevents indefinite pull-down if the fire button is held too long.
RCS Mode Settings
Determines when and how recoil control is applied.
RCS Activation Mode
Dropdown
See options below.
Enemy Detection Threshold
Slider (0–100%)
Only visible in "Enemy Detected" mode. Defines how centered the crosshair must be on the target for RCS to engage. Lower = stricter (must be dead center). Higher = more lenient.
Activation Mode options:
Always Active
RCS runs whenever the fire key is held, regardless of whether a target is present.
Enemy Detected
RCS activates only when a valid enemy is identified by the AI.
Within FOV
RCS activates only when an enemy is inside the defined FOV radius.
Pattern Based
Disables generic pull-down. Uses a specific CSV pattern file instead (enables Pattern Settings below).
Pattern Settings
Visible only when RCS Activation Mode is set to Pattern Based. Configures file-based recoil compensation that follows specific weapon spray patterns rather than just pulling down.
RCS Pattern File
Text field + Browse
Select a .csv file containing relative X/Y coordinates and timing data for a specific weapon's recoil pattern.
Pattern Randomness
Slider
Injects random pixel noise into the pattern to prevent perfectly static, robotic movement.
Pattern Interpolation
Slider
Adds intermediate smoothing points between pattern steps, creating fluid motion instead of jumping between coordinates.
Pattern Sensitivity
Slider
Scales the entire pattern's magnitude. Adjusts a single pattern file to work with different in-game sensitivities or FOV settings.
Require Enemy Visible
Checkbox
When enabled, pattern execution pauses if no enemy is detected. Prevents the cursor from moving erratically when firing at walls.
Pattern CSV Format
Pattern files are plain CSV with no header row. Each row represents one bullet in the spray sequence.
dx
Float
Horizontal pixel offset for this bullet (incremental, not cumulative). Positive = right, negative = left.
dy
Float
Vertical pixel offset for this bullet (incremental, not cumulative). The system negates this value at runtime, so a negative dy in the CSV produces a downward pull.
delay_ms
Float/Int
Milliseconds to wait before advancing to the next bullet step. Typically matches the weapon's fire rate.
Example — AK-47 pattern (first 6 bullets):
Example — simple vertical pull test:
How values are scaled:
CSV values go through two stages of scaling before reaching the mouse:
Load time:
value × 2.45 / game_sensitivity— the 2.45 factor converts from artanis-rcs units.Runtime:
scaled_value × Pattern Sensitivity slider— the UI slider (0.1–2.0) applies an additional multiplier.
Use the Pattern Sensitivity slider to adjust a single CSV file to work across different in-game sensitivities without editing the file.
Behavior notes:
Empty lines are skipped.
Lines with fewer than 3 columns are skipped.
Invalid numbers are skipped (no crash).
If the spray lasts longer than the pattern, the last row repeats indefinitely.
Files are read with UTF-8 encoding (BOM-safe).
Pattern files are stored in the
rcs_patterns/folder.
Pattern Visualization
Visible only in Pattern Based mode. Provides real-time visual feedback on the loaded pattern.
Full Pattern View
Graph showing the complete static path of the loaded CSV pattern.
Live Pattern View
Dynamic graph showing cursor position along the pattern path in real-time while firing.
Bullets
Total number of steps in the loaded pattern.
Current
Step index currently being executed.
Active
Whether the RCS system is currently engaging (Yes/No).
Reload Pattern
Forces a reload of the CSV file from disk. Useful after editing the file externally.
Humanizer Page
The Humanizer page masks robotic movement patterns to make aim behavior indistinguishable from natural human input. It uses a tab system with independent configurations for Aim and Flick behaviors.
Humanizer Tabs
Aim Humanizer
Standard tracking/smoothing (Aim V2 / Legacy)
Flick Humanizer
Rapid target acquisition (Flick / Trigger)
Each tab contains identical controls described below. Tune them independently — tracking and flicking have very different movement characteristics.
Humanization Mode
Enable
Checkbox
Master toggle for humanization on this tab (Aim or Flick).
Mode
Radio buttons
Traditional Humanization — software-based algorithms (WindMouse, Momentum, Entropy). MAKU Hardware Humanization — offloads to hardware emulation using geometric curves.
MAKU Hardware Controls
Note: Not available until a future MAKCU firmware update.
Visible only when "MAKU Hardware Humanization" is selected. Uses geometric curves to create fluid, non-linear paths.
Movement Type
Dropdown
Straight (2-param) — direct movement. Segmented (3-param) — multiple linear segments. Quadratic Bézier (5-param) — one control point, smooth simple curve. Cubic Bézier (7-param) — two control points, complex S-curves or arcs.
Segments
Slider
Resolution/smoothness of the curve. More segments = smoother path but more processing.
CX1 / CY1
Sliders
Coordinates for the first curve control point.
CX2 / CY2
Sliders
Coordinates for the second control point (Cubic mode only).
Traditional Humanization Controls
Visible only when "Traditional Humanization" is selected. Divided into several specialized subsystems.
WindMouse Algorithm
Simulates mouse movement as a physical object affected by gravity and wind forces.
Enable WindMouse
Checkbox
Toggles this specific algorithm.
Gravity (Min/Max)
Sliders
Force pulling the cursor toward the target. Lower = wider, arcing paths. Higher = snappier, straighter lines.
Wind (Min/Max)
Sliders
Chaotic side-to-side force that adds wobble/imperfection to the path.
Speed (Min/Max)
Sliders
Velocity range of movement steps.
Damp (Min/Max)
Sliders
Braking force that reduces wind as the cursor approaches the target. Ensures the cursor settles onto the target instead of oscillating.
Advanced Humanization (Always-On)
Momentum Tracking
Checkbox
Simulates physical mouse weight. Adds acceleration and deceleration phases so movement doesn't start/stop instantly.
Movement Clamping
Checkbox
Limits maximum pixels moved in a single frame (typically ±14px). Prevents superhuman instantaneous snaps.
Stop/Pause System
Randomly introduces micro-pauses to simulate repositioning or hesitation.
Chance
Slider (%)
Probability of a pause occurring per movement cycle.
Min Pause
Slider (ms)
Minimum pause duration.
Max Pause
Slider (ms)
Maximum pause duration.
Pattern Masking
Adds tiny, imperceptible jitter to break up mathematical linearity in small movements.
Intensity
Slider
Strength of the micro-adjustments.
Scale
Slider
Global multiplier for all humanization variance.
Sub-Movement Decomposition
Breaks long movements into smaller chunks separated by micro-delays, rather than one continuous slide.
Enable
Checkbox
Toggles sub-movement decomposition.
Chunk Size
Slider
Size of each sub-movement segment.
Proximity Pause
Triggers hesitation when the crosshair gets very close to the target, simulating a user confirming their aim.
Threshold
Slider (px)
Distance from target to trigger the pause.
Cooldown
Slider (ms)
Minimum time before this can trigger again.
Easing System
Controls acceleration curves for starting and stopping movements around pauses.
Ease-Out
Slider (ms)
Deceleration duration before a pause.
Ease-In
Slider (ms)
Acceleration duration after a pause.
Curve
Slider
Power of the easing curve. 1.0 = Linear, 2.0 = Quadratic, 3.0 = Cubic. Higher values create more pronounced acceleration/deceleration.
Momentum Tracking System (Physics-Based)
Fine-tuning for the momentum physics engine.
Decay
Slider
How fast momentum bleeds off (friction). Higher = stops faster.
Lead Bias
Slider
Forward prediction multiplier. Higher values make the system aim slightly ahead of where it thinks the target is going.
Deadzone
Slider (px)
Minimum error required before momentum corrections are applied. Prevents micro-corrections on already-accurate aim.
Correction Strength
Slider
How aggressively the system corrects overshoot/undershoot errors.
Entropy-Aware Humanization
Features specifically designed to defeat statistical analysis detection.
Speed Variance
Slider
Modulates movement speed with smooth noise to increase the Speed CV (Coefficient of Variation) metric. Makes speed less constant and more human-like.
Path Curvature
Slider
Adds sinusoidal waviness perpendicular to the movement path. Prevents perfectly straight aim lines.
Endpoint Settling
Slider
Adds micro-oscillations at the target point instead of stopping instantly. Simulates a human stabilizing their hand after a flick.
Tip: Use the Entropy Analyzer (Settings → Mouse/HID) to see real-time metrics while tuning these values.
Settings Page
The Settings page is divided into multiple sub-tabs covering hardware, detection, filtering, tracking, input devices, configuration management, and ML model training.
Capture Tab
Handles video input, image acquisition, and hardware acceleration.
Core Capture Settings
Capture Mode
Dropdown
Capture Card — physical hardware device. NDI / NDI-V2 — network video (dual-PC setups). Screenshot — Windows API screen grab (slow, for testing). UDP — custom UDP socket (currently non-functional).
Enable Crop
Checkbox
Crops the input to a center region instead of processing the full frame. Significantly improves performance.
Debug Size
Text field + Apply
Sets the size of the debug overlay window.
Select Visual Type
Dropdown
Debug - Crop Only — shows the cropped processing region with detections. Debug - Full Screen — shows the entire captured frame. Debug - Weapon Detection ROI — specialized view for OCR. None — disables visual feedback for maximum performance.
Debug Performance
Display Scale
Slider (0.25–1.0)
Downscales the debug window resolution. 0.5 renders at 50% resolution. Does not affect detection, only the viewer.
Max FPS
Slider (15–240)
Limits debug window refresh rate. Set to 30–60 to save resources. This does not limit detection FPS.
Hardware Source Settings
Visible only when Capture Mode is set to Capture Card.
Select Codec
Dropdown
Windows API for the camera interface. Direct Show — more compatible. MSMF — newer but generally worse compatibility.
Capture Source
Dropdown
Lists all connected video input devices (e.g., "Cam Link 4K").
Capture Format
Dropdown
Raw data format: YUY2, NV12, MJPG, etc. Match this to your capture card's supported formats.
Capture Resolution
Text fields (W × H)
Manually sets capture resolution (e.g., 1920 × 1080).
Max Capture FPS
Text field
Target framerate for the capture loop (e.g., 240). Prevents the capture thread from consuming 100% of a CPU core.
UDP Configuration
Note: UDP capture mode is currently non-functional.
Visible only when Capture Mode is set to UDP.
UDP URL
Text field
Network address to listen on (e.g., udp://127.0.0.1:1234).
Use MJPEG UDP
Checkbox
Switches decoding from H.264 to MJPEG.
H.264 Decode Method
Dropdown
Hardware decoder: Auto, NVDEC (NVIDIA), QSV (Intel), AMF (AMD), CPU. Match to your GPU for lowest latency.
Frame Dropping
Dropdown
Disabled (Smooth) — processes all frames. Balanced — moderate dropping. Aggressive (Low Latency) — drops old frames to always process the newest data.
Debug Capture
Checkbox
Enables verbose logging for diagnosing stream issues.
OpenCL GPU Acceleration
Enable OpenCL
Checkbox
Offloads color space conversion (YUV → RGB) from CPU to GPU.
Performance Mode
Dropdown
Force CPU — maximum inference speed (avoids PCIe overhead). Auto — uses GPU if efficient. Force GPU — strictly uses GPU acceleration.
Early Initialization
Checkbox
Pre-loads OpenCL contexts at startup. Increases launch time but prevents lag spikes when detection first starts.
Status
Label
Displays current acceleration state (e.g., "GPU Acceleration Active").
Detection / Model Tab
Configures the neural network and inference engine that powers object detection.
Backend & Model
Backend
Dropdown
Inference framework. Defaults to ONNX (Open Neural Network Exchange).
ONNX Provider
Dropdown
gpu — uses DirectML (Windows). openvino — uses Intel OpenVINO (works for AMD CPUs too).
OpenVINO Device
Dropdown
Only visible with openvino provider. Select hardware: AUTO, CPU, GPU, NPU.
ONNX Threads
Dropdown
CPU threads dedicated to ONNX runtime. Tuning this balances CPU load vs. detection speed.
Model File
Dropdown
The AI model to load (.onnx, .enc, or .xml files from models/). Determines what the system detects.
Confidence
Dropdown (1–99%)
Minimum confidence threshold for a detection to be valid. Higher = fewer false positives but may miss targets. Lower = catches more but risks false detections.
Model Settings
Model Type
Dropdown
Object Detection — standard bounding boxes. Pose Estimation — keypoint detection (head, neck, shoulders, etc.). Requires a pose model.
Async Inference Enable
Checkbox
Enabled — runs inference in parallel with capture. Higher FPS but slightly more latency. Disabled — synchronous, lowest latency (frame-by-frame).
Target Keypoints
Dropdown
Only visible in Pose Estimation mode. Select body part to target: Head, Neck, Chest, or All Keypoints.
Detection Filtering
Bone Selection
Dropdown
Nearest Bone (Simple/Advanced) — closest bone to crosshair. Head/Chest — forced area targeting. Pose Estimation — uses keypoint data.
Class Filter 0–3
Checkboxes
Enable specific model class IDs to prioritize one class over another.
Disable Class 0–3
Checkboxes
Explicitly ignore specific class IDs. The system will not lock onto disabled classes even if detected.
Filtering Tab
Advanced logic to ignore specific targets or false positives using computer vision techniques.
Enemy Color Filter
High-level filter for games with character outlines or team colors.
Enemy Color Filter Enable
Checkbox
Toggles color-based validation.
Color Selection
Dropdown
Color range to look for on targets: Purple, Pink, Yellow, Red (with variants). Targets without this color are ignored.
Color + Shape Ignore Filter
Define rules to ignore detections based on geometric properties or internal colors. Useful for filtering teammates, dead bodies, or map objects.
Enable
Checkbox
Master toggle for shape filtering.
Show Debug Overlay
Checkbox
Draws visualization on the debug screen showing search regions and detected shapes.
Decay
Spinbox (ms)
How long a target remains flagged as "ignore" once detected. Prevents flickering between valid/invalid states.
Filter Rules List
List + Add/Edit/Remove
Each rule can combine: Color (RGB + tolerance), Shape (Triangle, Circle, Diamond, Arrow), Rotation (specific angles), Dimensions (min/max area, aspect ratio), Region (Inside, Above, Below, Custom Offset relative to the bounding box).
Template Matching Ignore Filter
Ignores targets if a specific image (template) is found on or near them. Highly effective for filtering UI elements like "Revive" icons, nameplates, or status markers.
Enable
Checkbox
Master toggle for template matching.
Template Rules List
List + Add/Edit/Remove
Each rule includes the settings below.
Per-rule settings:
Template Image
The .png or .jpg icon to search for.
Preprocessing Pipeline
Basic: Grayscale, Invert, Crop. Enhancement: CLAHE (contrast), Sharpen, Bilateral Filter. Noise Removal: Gaussian Blur, Morphology (Erode/Dilate), Small Blob removal. Edge Detection: Canny edge extraction for shape-based matching.
Threshold
How close the match must be (0–100%).
Method
Matching algorithm: Correlation, Squared Difference, etc.
Multi-scale
Checks multiple sizes of the template to handle distance scaling.
Search Region
Where to look relative to the detection box (e.g., "Above BBox" for health bar icons).
Live Preview
Split-screen view showing the original image and processed result to verify settings work.
Tracking Tab
Controls how aim assist interprets target movement, stabilizes aiming, and predicts future position.
Basic Tracking
Min Smoothing Clamp
Slider
Hard lower limit on smoothing values. Prevents dynamic algorithms from reducing smoothing to near zero, which would cause wild flicks.
Tracking Deadzone
Slider (px)
Small radius around the target center where no adjustments are made. Prevents jitter when the crosshair is already on target.
Extra Smoothing
Slider
Additional uniform smoothing layer on top of the base calculation. Useful for games with erratic mouse input.
Recommended starting values: Min Smoothing Clamp 3–5. Tracking Deadzone 2–4 px. Extra Smoothing 0–2.
Prediction System
The primary predictive aiming system. Uses a Kalman filter to estimate where a moving target will be in the future, compensating for system latency. All internal filter parameters (process noise, measurement noise, velocity smoothing, etc.) are derived automatically from the Responsiveness meta-slider.
Enable Predictive Aim
Checkbox
Master toggle for the prediction system. When disabled, no lead/prediction is applied.
Show Kalman Debug Info
Checkbox
Prints Kalman filter state (position, velocity, covariance) to the console. Only visible when prediction is enabled.
Preset
Dropdown
Quick-apply a prediction tuning profile. Manually adjusting any slider switches to "Custom."
Sliders:
Aim Ahead (ms)
0–200 ms
How far ahead to lead the target. 0 = no lead, 100 = 100ms lead. This is the single most important prediction parameter — it directly controls how much the system compensates for your total pipeline latency.
Responsiveness
0–200
Single meta-slider controlling all internal filter parameters. 0 = Smooth (slow tracking, heavy smoothing). 100 = Snappy (fast tracking, minimal smoothing). 200 = Ultra-Snappy (most aggressive tracking). This drives process noise, measurement noise, velocity smoothing, and other values automatically.
Lock Delay (ms)
0–500 ms
How long to track a target before prediction kicks in. Longer = more cautious startup. Shorter = faster prediction. 0 = instant prediction with no warmup.
Tracking Zone (px)
5–50 px
How close the crosshair must be to the target for lock to activate. Smaller = stricter lock requirement. Larger = more forgiving acquisition.
Y Prediction (%)
0–100%
How much vertical prediction to apply. 0% = no vertical prediction. 100% = full vertical prediction. Lower values prevent upward drift toward the head on horizontally-moving targets.
Optional:
Velocity Aim Ahead
Checkbox + Slider (0–200%)
Scales aim-ahead time by target velocity. Faster targets get more lead, slower targets get less. The slider controls intensity — 100% = standard scaling, 200% = aggressive scaling.
Tuning guide: Start with Aim Ahead at 5ms and Responsiveness at 45. If tracking feels laggy, increase Responsiveness. If aim leads too far ahead of the target, reduce Aim Ahead. Set Lock Delay to 100–200ms to prevent prediction from engaging on brief flick-overs. Keep Y Prediction at 20–30% unless targets move vertically often.
Latency & Prediction Debug Info
Real-time performance readout panel.
Capture Latency
Time for the image to reach the software.
Processing Latency
Time for AI inference and logic.
Total Latency
Sum of all delays in the pipeline.
Velocity
Current calculated target velocity (X/Y).
Prediction Status
Whether prediction is active or warming up.
Samples
Number of valid data points collected for the current track.
Mouse / HID Tab
Hardware configuration center. Manages how the software communicates with your mouse, handles input blocking, and normalizes sensitivity.
Device Selection
Device
Dropdown
Moonlink — software-based mouse simulation, no hardware needed. KMBox — hardware mouse spoofing device. MAKCU / Ferrum — advanced hardware spoofing devices.
COM Port
Dropdown
Only visible for hardware devices. Select the USB port your device is connected to.
Hardware setup: After connecting your device, select the correct COM Port from the dropdown. If the device isn't listed, check Device Manager to confirm the port number and that drivers are installed.
Input Masking
Intercepts or blocks physical mouse clicks to prevent conflict between your input and the AI.
Mouse 1 / 2 / 4 / 5
Checkboxes
When checked, the software blocks the physical input of that button while the AI is active. Prevents "fighting" the system.
M1 Block Threshold
Slider (%)
Only visible when Mouse 1 is checked. Defines a safety zone as a percentage of the target's bounding box width. Lower = stricter, aim must be very close to center before your click registers. Higher = more lenient, allows firing near the edge.
Debug & Analysis
Show Button States Debug
Checkbox
Only visible for MAKCU/Ferrum devices. Displays internal hardware button state to verify input masking works on the physical device.
Enable Entropy Analyzer
Checkbox
Opens the Entropy Analyzer panel. See Entropy Analyzer for details.
Sensitivity Normalization
Ensures aim settings feel consistent regardless of mouse DPI or in-game sensitivity.
Enable Sensitivity Normalization
Checkbox
Toggles normalization. When enabled, Speed 10 feels the same at 800 DPI as at 1600 DPI.
DPI Value
Input field
Your physical mouse DPI (e.g., 800, 1600).
In-Game Sens
Input field
Your actual in-game sensitivity value.
Reference Sens
Input field
Baseline value used by config creators. The software calculates a multiplier to match the creator's feel on your setup.
Legacy Distance Smoothing
Checkbox
Disabled (Recommended) — newer logic: aim is slower/precise at close range, faster at long range. Enabled — old behavior: slower at range, faster up close.
Cross-reference: The Aim V2 Distance Scale Factor slider only functions when Legacy Distance Smoothing is enabled.
Delays & Sensitivity
Device Delay
Slider (ms)
Delay added to hardware device commands. Increase to fix connection instability or "ghosty" movement.
Aim Delay
Slider (ms)
Artificial latency added to Aim commands.
Flick Delay
Slider (ms)
Artificial latency added to Flick commands.
RCS Delay
Slider (ms)
Artificial latency added to Recoil commands.
Sens Multiplier
Slider
Global multiplier for all mouse output. 1.0 = standard. < 1.0 = reduces speed. > 1.0 = amplifies speed.
Config Tab
Administrative hub for file management, profile switching, and automatic weapon detection. Divided into three sub-tabs: Settings, OCR, and Template Matching.
Settings Sub-Tab
Profile Management
Config File
Dropdown
Select the active JSON configuration file.
Refresh List
Button
Reloads the file list from the configs/ folder.
Import Config
Button
Download configurations from a linked Discord channel.
Export Config
Button
Upload your current config and AI model to Discord for sharing.
Duplicate Config
Button
Creates a copy of the current config with a new name.
Delete Config
Button
Permanently removes the selected config file.
Weapon Class Management
Weapon Class
Dropdown
Manually switches the active profile. All settings on Aim, Flick, and RCS pages update instantly.
Manage
Button
Opens a window to create, rename, or delete weapon classes.
Hotkey
Input field + Set/Clear
Bind a key (e.g., F3) to cycle through weapon classes in-game.
OCR Sub-Tab (Text Recognition)
Uses Optical Character Recognition to read text on screen (e.g., weapon name in the HUD) and automatically switch Weapon Class.
OCR Engine
Enable Auto Weapon Detection
Checkbox
Master switch for text recognition.
OCR Engine
Dropdown
Recognition library (default: EasyOCR).
ROI Configuration (Region of Interest)
Defines where on screen the system looks for text.
ROI Position (X / Y)
Sliders
Top-left corner coordinates of the search box.
ROI Size (W / H)
Sliders
Width and height of the search box.
Live Preview
Start/Stop Preview
Button
Shows a real-time feed of the capture area.
Green Box
Visual
Indicates the current scan region.
Zoom
Dropdown
Zoom into specific corners/center to position the ROI precisely over the weapon name.
Refresh/Scale
Sliders
Adjust preview framerate and quality.
Preprocessing & Detection
Grayscale / Threshold
Sliders
Filters to make text stand out against the game background.
Contrast
Slider
Separation between text and background.
Detection Interval
Slider (ms)
How often the system scans. Higher values save CPU.
Similarity Threshold
Slider (0–100%)
How close detected text must match your keyword. Lower allows fuzzy matching (e.g., "Vanda1" → "Vandal").
OCR Languages
Dropdown
Language packs for reading text.
Weapon Text Mappings
Mappings Table
List
Keyword — text to look for (e.g., "AK-47"). Target Class — profile to switch to (e.g., "Rifles").
Test OCR Mapping
Button
Type a phrase to test if it triggers a match against current rules.
Template Matching Sub-Tab (Icon Recognition)
Alternative to OCR. Looks for specific images (weapon icons in HUD/killfeed) to trigger profile switches.
Template Settings
Enable Template Matching
Checkbox
Master switch for icon recognition.
Check Interval
Slider (ms)
How frequently the system scans for images.
Switch Cooldown
Slider (ms)
Minimum wait before switching classes again. Prevents rapid flickering.
Live Debug Status
Status
Active state and number of loaded rules.
Last Match
Name of last detected icon and confidence score.
Checks/Matches
Activity counters.
Template Rules
Add Rule
Button
Opens dialog: Template Image (browse for .png/.jpg), Target Weapon Class, Search Region (e.g., "Bottom Right"), Match Threshold (0.80+ recommended), Preprocessing (Blur, Edges, Invert).
Edit/Remove
Buttons
Modify or delete existing rules.
Test Selected
Button
Verifies if the selected rule matches the current screen.
ML Mouse Model Tab
Train a custom neural network that mimics your specific mouse movement style. Instead of generic humanization algorithms, this system learns from your actual hand movements.
Training Data Collection
Record your natural mouse movements to build a dataset.
Status
Label
Current state: "Not Recording" or "Recording - Move your mouse naturally!"
Sensitivity
Slider
Match to your in-game sensitivity so the AI learns correct speed/distance relationships.
Start Recording
Button
Begins capturing mouse data. Move your mouse naturally — flick, track, stop — as if aiming at targets.
Stop Recording
Button
Saves the captured session to the database.
Live Stats
Label
Real-time feedback: movements captured, current speed, movement classification.
Tip: Record at least 5–10 sessions of 30+ seconds each. Vary your movements — track slow targets, flick to distant ones, make small corrections. The more diverse your data, the better the model.
Training Data Summary
Total Sessions
Number of saved recording sessions.
Recent Sessions
Details on the last few recordings.
Metrics
Path efficiency (straightness), jitter (shakiness), overshoots.
Refresh Summary
Updates stats with the latest data.
Model Training
Model Architecture
Dropdown
VAE (Original) — balanced quality/performance. GRU (Fast) — faster processing, ideal for older PCs. Transformer (Best) — highest quality, recommended for high-end PCs.
Training Epochs
Spinbox
How many passes the AI makes over your data. 200–500 recommended. Higher = more accurate but longer training.
Train Model from Data
Button
Starts training. Generates a .pth or .onnx model file.
Load Existing Model
Button
Re-loads the last used model into memory.
Model Info
Label
Architecture type and Validation Loss (lower = more accurate mimicry).
Model Management
Available Models
Dropdown
Lists all trained models with name, type, and accuracy score.
Load Selected
Button
Activates the chosen model.
Save As
Button
Creates a copy with a new name.
Rename
Button
Changes the model's filename.
Delete
Button
Permanently removes the model file.
Data Management
Export Training Data
Button
Saves raw movement dataset to a file for backup or sharing.
Clear All Data
Button
Wipes all recorded sessions. Use when starting fresh with a different aiming style.
Entropy Analyzer
The Entropy Analyzer is a real-time diagnostic panel that measures the statistical "humanness" of your mouse output. It is enabled via the checkbox in Settings → Mouse/HID → Debug & Analysis.
What It Measures
Speed CV
Coefficient of Variation of movement speed. Low values indicate constant, robotic speed.
0.4–0.8
Path Linearity
How straight the cursor path is. 1.0 = perfectly straight (robotic).
0.6–0.85
Acceleration Entropy
Randomness in acceleration changes. Low = predictable ramp-up/down.
> 2.0
Pause Distribution
How natural the pattern of micro-pauses is.
Varies
Direction Change Rate
Frequency of small direction adjustments during movement.
> 0.1
How to Use It
Enable the Entropy Analyzer.
Start an aim session and observe the real-time metrics.
If any metric is flagged red or falls outside the target range, adjust the corresponding Humanizer setting:
Low Speed CV → increase Speed Variance (Entropy-Aware Humanization).
High Path Linearity → increase Path Curvature.
Low Acceleration Entropy → increase Momentum Tracking decay or enable WindMouse.
No pauses → enable Stop/Pause System and increase Chance.
Re-test until all metrics fall within acceptable ranges.
Debug Overlay
When a visual debug type is selected in Settings → Capture → Select Visual Type, an overlay window shows real-time detection information.
Visual Elements
Bounding Boxes
Green/Red
Detected targets. Green = valid target. Red = filtered/ignored.
FOV Circle
White/Blue
The current aim FOV search area, centered on your crosshair.
Prediction Vector
Yellow arrow
Direction and magnitude of Kalman prediction (where the system thinks the target is heading).
Confidence Label
White text
Detection confidence percentage displayed above each bounding box.
Keypoint Skeleton
Colored dots/lines
Visible in Pose Estimation mode. Shows detected body joints and connections.
Filter Debug
Red markers
When Shape/Template filter debug overlay is enabled, shows matched regions and filter decisions.
Crosshair
Small cross
Center reference point (where your aim is).
Debug View Modes
Debug - Crop Only
Only the cropped processing region, at full detail. Best for tuning aim and detection.
Debug - Full Screen
The entire captured frame with overlays. Useful for verifying crop position.
Debug - Weapon Detection ROI
The OCR region with preprocessing visualization. Use when tuning OCR settings.
None
No overlay. Maximum performance. Use during actual gameplay.
Hotkeys Reference
All bindable keys in one place.
Aim Key
Aim Page
Activates aim assist while held/toggled.
Flick Key
Flick & Trigger Page
Activates flick assist.
Trigger Key
Flick & Trigger Page
Activates triggerbot while held.
Weapon Class Cycle
Settings → Config → Weapon Class
Cycles to the next weapon class profile.
Start/Stop
Sidebar
Starts or stops the detection loop.
Reset Session
Sidebar
Zeroes all runtime state (EMAs, PID, Kalman, serial lock).
Plugins
The Plugins tab provides a Lua-based extension system for adding custom functionality.
Managing Plugins
Plugin List
Shows all discovered plugins with enable/disable toggles.
Refresh
Re-scans the plugins directory for new files.
Enable/Disable
Toggle individual plugins on or off. State is saved to your config.
Config File Format
MoonLightAI stores settings in JSON configuration files located in the configs/ folder.
Structure
Sections
DEFAULT
Global settings (capture, tracking, mouse/HID, delays)
No
AIMBOT
Aim, Flick, RCS, and Humanizer settings
Yes
PRESETS
Saved preset profiles
No
WEAPON_CLASSES
Weapon class definitions and names
No
Key Behaviors
Auto-defaults: Missing keys are automatically added with default values on startup. You can safely delete a key from the file and it will be restored.
Type validation: Invalid value types are auto-corrected to defaults on startup (e.g., a string in a float field is replaced with the default float).
Live reload: Changes made in the UI are saved to the config file immediately.
File watcher: External changes to the config file are detected and reloaded automatically.
Switching configs: Selecting a different config from the dropdown reloads the entire UI and syncs all runtime values.
Manual Editing
You can edit config files directly with a text editor. Changes are picked up by the file watcher. All values are stored as strings (requirement of the config parser):
Glossary
Bézier Curve
A mathematical curve defined by control points. Used in MAKU humanization to create smooth, non-linear mouse paths.
Bounding Box
A rectangle drawn around a detected target by the AI model. Defined by X, Y, width, and height.
Confidence
A 0–100% score from the AI model indicating how certain it is that a detection is valid.
CV (Coefficient of Variation)
A statistical measure of speed variability. Low CV = constant speed (robotic). High CV = variable speed (human).
Deadzone
A small region where no adjustments are made. Prevents jitter when already on target.
DirectML
Microsoft's hardware-accelerated machine learning API. Used as the GPU backend for ONNX inference on Windows.
EMA (Exponential Moving Average)
A smoothing technique that gives more weight to recent values. Used for velocity smoothing and signal filtering.
Entropy
A measure of randomness/unpredictability. Higher entropy in mouse movements indicates more human-like behavior.
FOV (Field of View)
The circular search area around the crosshair. Only targets within this radius are considered.
Interpolation
Adding intermediate points between defined steps to create smooth transitions.
Kalman Filter
A mathematical algorithm that predicts future state (position, velocity) from noisy measurements. Used for target prediction.
Keypoint
A specific body joint detected by pose estimation models (e.g., head, neck, shoulder, hip).
Latency
Total delay from screen capture to mouse movement output. Lower is better.
NDI (Network Device Interface)
A protocol for sending video over a local network. Used in dual-PC setups.
ONNX (Open Neural Network Exchange)
A standard format for AI models. Provides broad compatibility across hardware.
OpenCL
A framework for GPU-accelerated computation. Used here for image color conversion.
OpenVINO
Intel's inference toolkit. Optimized for Intel CPUs and integrated GPUs. Also works on AMD CPUs.
Overshoot
When the cursor moves past the target before correcting back. Some amount is natural for humans.
PID
Proportional-Integral-Derivative controller. A feedback loop used in some motion algorithms.
Pose Estimation
An AI technique that detects specific body joints/bones rather than just bounding boxes.
Process Noise (Q)
Kalman filter parameter. Controls expected acceleration/velocity changes. Higher = more responsive to direction changes.
Measurement Noise (R)
Kalman filter parameter. Controls trust in raw detection data. Higher = smoother but slower to react.
ROI (Region of Interest)
A defined rectangular area on screen where OCR or template matching searches for content.
Smoothing
A value that controls how gradually the cursor moves toward the target. Higher = slower/smoother. Lower = faster/snappier.
Weapon Class
A named profile within a config file (e.g., "Rifles," "Snipers") with its own aim, flick, RCS, and humanizer settings.
WindMouse
A humanization algorithm that simulates mouse movement as a physical object affected by gravity and wind forces.
Troubleshooting
Capture Issues
Capture card not detected
Verify the device appears in Windows Device Manager under "Cameras" or "Imaging Devices."
Try switching between Direct Show and MSMF codecs.
Ensure no other application (OBS, Discord) is using the capture device.
Check that the correct resolution and format are set.
NDI source not appearing
Confirm both PCs are on the same local network.
Verify the NDI source is actively broadcasting (check NDI Monitor on the sending PC).
Temporarily disable firewalls to test connectivity.
Low capture FPS
Enable Crop to reduce the processed area.
Lower Capture Resolution if your card supports it.
Check Max Capture FPS is set appropriately (not too low).
Ensure OpenCL is enabled for GPU-accelerated color conversion.
Detection Issues
No detections appearing
Verify a model file is loaded (Settings → Detection/Model).
Lower the Confidence threshold (try 30–40% for testing).
Check that the correct Model Type is selected (Object Detection vs. Pose Estimation).
Ensure the debug view is enabled to see bounding boxes.
Too many false positives
Increase Confidence threshold.
Enable and configure Filtering (color, shape, or template matching).
Use Class Filtering to disable irrelevant class IDs.
Detection FPS is low
Enable Async Inference.
Reduce crop size.
Switch ONNX Provider to
gpuif currently on CPU.Reduce ONNX Threads if CPU is overloaded.
Use a smaller/faster model file.
Aim Issues
Aim feels "floaty" or slow
Decrease Smooth X/Y values.
Reduce Extra Smoothing on the Tracking tab.
Check that Sensitivity Normalization values match your actual DPI/sens.
Verify Sens Multiplier is at 1.0.
Aim is jittery or twitchy
Increase Tracking Deadzone.
Increase Kalman Measurement Noise (R).
Increase Smooth X/Y values.
Enable Momentum Tracking in Humanizer.
Aim overshoots targets
Increase smoothing values.
Reduce Kalman Prediction Responsiveness.
Enable Humanizer → Momentum Tracking with higher Decay.
Enable Endpoint Settling in Entropy-Aware Humanization.
Config not updating when switching files
The dropdown should reload all UI widgets and runtime values automatically.
If values seem stale, click Refresh List and reselect the config.
Check the console output for error messages during config loading.
Hardware Device Issues
KMBox/MAKCU/Ferrum not responding
Verify the correct COM Port is selected.
Check that the device appears in Device Manager under "Ports (COM & LPT)."
Try increasing Device Delay (start with 5–10 ms).
Unplug and reconnect the device.
Ensure no other software is using the same COM port.
"Ghosty" or stuttering mouse movement with hardware devices
Increase Device Delay incrementally (2–5 ms at a time).
Increase the specific Aim/Flick/RCS Delay sliders.
Check USB cable quality and port (use USB 2.0 ports for serial devices).
OCR Issues
OCR not reading weapon names
Use the Live Preview to verify the ROI is positioned exactly over the weapon text.
Adjust preprocessing (Grayscale, Threshold, Contrast) until the text is clearly visible in the preview.
Lower the Similarity Threshold to allow fuzzy matching.
Verify the correct OCR Language is selected.
Performance Issues
General lag or high CPU usage
Enable Crop (single biggest performance improvement).
Set Debug Visual to "None" during gameplay.
Lower Display Scale and Max FPS for the debug window.
Enable Async Inference.
Reduce ONNX Threads if the CPU is saturated.
Disable OpenCL Early Initialization if startup is slow (does not affect runtime).
Last updated