Drop an image or Choose image file Loaded images stay in this browser; the app does not upload them.

Training status

iteration 0 / 0

splats -

RGB L1 -

SSIM global / local p10 - / -

PSNR RGB -

image size -

low-alpha / outside - / -

speed / elapsed - / -

GPU active / reserved 0 / 0 MB

state idle

Image2SplatPaint

Paint one image with splats.

Preview

Algorithm Max image side Initial splats Max splats Iterations Preview refresh

Curved Brush settings

Paint with curved Brush Splats using linked strokes or learned internal bends.

Training path Stroke texture Scale-matched repaint Stroke optimization Coarse-to-fine paint curriculum Fixed stroke opacity Starting width (long side /) Starting length (%) Residual move / stage (px) Initial width min (%) Initial width max (%) Front width max (%) Front width learning (%) Color anchor Stroke width (%) Edge-guided fine strokes

Splats per linked stroke Min Max

Linked-stroke learning (%)

Linked-stroke width profile Ends (%) Center (%)

Alternating stroke-structure updates (experimental)

Structure-focused update Interval Image anchor (%)

Internal control points Control-point positions (%) Splat size variation (%) Movement limit (px) Curved Brush underpainting Underpainting from P1 (experimental) Underpainting budget (%) Underpainting size variation (%) Maximum ribbon length (%)

Shared training settings

Training feedback Update quality metrics during training

Device memory Allows a much larger working set on high-memory devices. Device or browser termination is more likely.

Color workflow Monochrome underpainting Color finish starts at (%)

Layer and visibility

Core layer processing Rectangle and Brush keep these layer operations enabled as part of their paint training.

Tile/subtile culling Train layer order Layer-aware accumulation Discrete layers

Layer update interval Layer count Layer move radius

Optional size-based reorder Scale-biased Surface Layer Prior

Size-sort layer count Apply until (%) P1 sort interval P2 sort interval P3 sort interval

Optional contribution cleanup Current-Visibility Compaction v2

Contribution compaction starts at (%) Contribution compaction interval Contribution removal cap (%) Near-zero removal cap (%) Contribution window (iterations)

Optimizer

Position Color Opacity Alpha loss weight Scale Rotation Theta align Max anisotropy Max planar scale Edge containment σ

Growth and detail

Adaptive placement (%) Detail coherence

Phase-relative scale-spread guard

P1 minimum / phase median P2 minimum / phase median P3 minimum / phase median

Stage-aware growth

P1 growth share % P2 growth share % P3 growth share % Splat growth interval Growth per event % Growth apply until (%) Growth signal threshold

Structure-guided splat allocation

Structure allocation grid

Mid-training over-density correction

Over-density correction schedule Over-density correction interval (iterations) Low-utility donor candidates (%)

Rectangle Splats settings

Splat shape

Paint appearance

Learned opacity (before gradient) Min Max

Directional opacity multiplier (short / long) Min Max

Center-to-edge opacity multiplier Min Max

Trapezoid edge ratio (top / base) Min Max

Aspect ratio (long side / short side) Min Max

Orientation Orientation tolerance (°) Preserve footprint area Edge-directed taper Structure-aware ratio Asymmetric edge softness

Illustrative Brush shape settings

Paint appearance

Learned opacity (before gradient) Min Max

Directional opacity multiplier Min Max

Center-to-edge opacity multiplier Min Max

Aspect ratio (long side / short side) Min Max

Train directional width taper Min Max

Local color-flow orientation Directional stroke aspect floor

Ribbon minimum Accent minimum

GS Virtual Camera Sampling

Bounded virtual depth 3DGS-style multi-view: one front or virtual camera supplies all trainable gradients each iteration. GOF density ranking Virtual camera share (%) Virtual camera max angle Virtual camera count Shared camera FOV Teacher coverage: load an image to estimate.

GPU memory details

GPU memory mode

limitchecking

estimate-

budget headroom-

limiter-

recommended-

speed-

detail cap-

Splat capacityManual limit

tile reserve-

measured speed-

Auto keeps browser and system headroom from standard memory hints. Exact free VRAM is unavailable.

PlayCanvas tilt check Engine 2.20.6

Show cameras used in this run

Finish training to inspect the cameras used in this run.

Camera framing Pitch X 0° Yaw Y 0°

Camera: center orbit
Position: -
Radius: -
FOV: 50°
Projection check: -

Virtual teacher images used in this run.

Finish training to inspect the PLY.

Finish training to inspect the splat parameters.

PLY uses standard SH0 Gaussian Splatting fields. For conventional multi-angle 3DGS viewing, train with GS Virtual Camera Sampling; Planar Gaussian is front-view only.

splats-

Current Splats preview, including shape, effects, alpha background, and optional outside-image padding.

PNG resolution Long side (px)

Training size

Finish training before export.