Recreate or stylize an image with trainable splats.
Runs directly in your WebGPU browser.
Launch the app · Quick start · Algorithms · Development
WebGPU Browser-only training MIT
Image2SplatPaint explores faithful image representation and differentiable
stylization with trainable splats. Planar Gaussian and
GS Virtual Camera Sampling focus on reproduction, while Rectangle Splats
and Brush Splats use geometric shapes, strokes, and independent paint layers
to create deliberately stylized results. For these stylization paths, visible
character matters alongside numerical fidelity. They can also act as trainable,
shape-aware blur filters that simplify fine detail into larger paint forms.
- Choose an Algorithm. Algorithm and setting changes apply to the next Train without reloading the page or clearing the loaded image and current result.
- Select Load image, use Sample, or drop an image on the canvas.
- Set Max image side, splat counts, and Iterations, then select Train. Use Pause or Stop if needed.
- Switch between Original and Splats, then review the visible result. RGB L1, SSIM, and PSNR help evaluate fidelity and compare runs; for Rectangle and Brush results, also evaluate their visible style and stroke structure. The status area reports GPU and runtime state.
- Open Export to save the current rendering as PNG. Gaussian results can also be saved as a standard 3DGS PLY; virtual-camera results can be inspected in Tilt.
Important
Training requires a WebGPU-capable browser and GPU. The maintained release target is current desktop Chrome on macOS.
Larger image and splat limits can use substantial GPU memory and take longer to finish. The page shows a visible explanation and retry action when WebGPU cannot be initialized.
Loaded images, training state, previews, and generated results are processed locally in the browser. The app has no image-upload or analytics endpoint. GitHub Pages still serves the static application files, and browser or hosting logs are outside the app's control. Training state is not persisted across a reload, so save PNG or PLY results you want to keep.
| Algorithm | Purpose | Exact export | Tilt |
|---|---|---|---|
Planar Gaussian |
Stable front-view approximation of one image | PNG and standard 3DGS PLY | No |
Rectangle Splats |
Analytic rectangle, trapezoid, or triangle paint shapes | PNG | No |
Brush Splats |
Directional illustrative brush shapes | PNG | No |
GS Virtual Camera Sampling |
Thin-depth 3DGS-style training from front and virtual teachers | PNG and standard 3DGS PLY | Yes |
All four choices share one custom WebGPU optimizer and standard front-to-back
alpha compositing. Each algorithm keeps its own initialization, kernel, opacity
semantics, settings, metrics, and export capability. Grid initialization and
density growth respect the source image's pixel aspect. GS Virtual Camera Sampling can additionally learn a small bounded depth and is the only path that
enables virtual-camera teachers and the Tilt tab.
The selector configures the next run. Once a result exists, Export eligibility and Tilt availability stay bound to that completed result until Reset, Clear, image replacement, or a new completed run.
Advanced Rectangle and Brush behavior
Rectangle Splats exposes Min and Max values for
Short edge / long edge. 1 / 1 keeps the current rectangular kernel.
Lower values taper the short parallel edge while the opposite edge stays at
full width: 0 allows triangle tips, and 0 <= Min <= Max <= 1 distributes a
deterministic mix of triangles, trapezoids, and rectangles across the paint
layers. Optional Rectangle-only controls can preserve each footprint area
while tapering, point the narrow edge toward stronger local structure, prefer
the selected Max ratio in flat regions, and use a harder narrow edge with a
softer wide edge. 1 / 1 retains the existing Rectangle training path.
Learned opacity (before gradient) bounds each Rectangle's trainable opacity
to 0.005...0.995. Opacity gradient multiplier (short / long) is a fixed
0...1 multiplier that changes linearly from the trapezoid's short edge to its
long edge. Final opacity is learned opacity × gradient multiplier. The defaults
0.995 / 0.995 and 1 / 1 preserve the former uniform 0.995 behavior.
Brush settings apply at the next Train start. Experimental checkboxes remain off by default. Equal directional endpoints are treated as a uniform/no-taper setting, so the defaults retain the accepted Brush path.
| Setting | What it changes |
|---|---|
Learned opacity (before gradient) |
Bounds each trainable Brush opacity to Min...Max in the safe 0.005...0.995 range. The default is 0.995 / 0.995. |
Opacity gradient multiplier |
A fixed 0...1 directional multiplier; it is not trained. Final opacity is learned opacity multiplied by this gradient in training, preview, standard-alpha overlap, and PNG export. The default 1 / 1 is uniform. |
Aspect ratio (long side / short side) |
Sets Brush-specific Min and Max anisotropy in one row. Defaults are 1 / 8; both values override the Shared Max anisotropy behavior during Brush training. |
Train directional width taper |
Learns a separate taper amount per splat between the configured directional Min and Max widths. Equal endpoints disable the directional change; the default 1 / 1 preserves the prior untapered path. |
Local color-flow orientation |
Softly aligns nearby directional splats when their colors and paint layers are similar. Broad patches and strong direction crossings are excluded. This is experimental. |
Directional stroke aspect floor |
Softly maintains a minimum long/short ratio while preserving footprint area. Ribbon minimum defaults to 2.2; Accent minimum defaults to 2.8; Base Patches are unchanged. These are lower bounds, while Maximum long / short ratio is the upper bound. This is experimental. |
The general Brush Min/Max applies to every Brush splat. When Directional stroke aspect floor is enabled, Ribbon and Accent additionally use their stronger family-specific minimums, capped by the Brush Max.
New Brush detail children always inherit their parent's paint layer. They can move later through the shared contribution-aware layer training; the former birth-time one-layer promotion has been removed. When layer training moves a Paint splat forward, stale RGB is repaired from its source-image footprint before the new order becomes visible.
The former training-teacher preprocessing, saturation gradient, Brush Line layer, Brush-profile choices, and rejected Sector-aware, optical-smoothing, and residual-matching experiments are not part of the product UI. Their comparison records remain in the local Brush experiment registry.
The app probes supported image headers before decoding. Large supported images
use a bounded decoder when available and are cached at no more than 4096 pixels
on the long side. Max image side is a separate training-time resize. The
status bar reports the current cached or training size, not the pre-cache source
dimensions. Decoder support varies by browser; the app retains a guarded
fallback for formats without bounded decode support. Training resize and GPU
estimates use the decoded display orientation, so EXIF-rotated images keep the
same aspect ratio when Train starts.
The status area reports progress, splat count, image quality, coverage, speed,
elapsed time, and tracked GPU use. Optional live quality updates are off by
default because full-image evaluation can slow training. The shared
Monochrome underpainting option begins with lightness and switches to RGB at
the selected point; final quality metrics always evaluate the RGB result.
- PNG exports the current rendered result for every algorithm.
- Splat PNG resolution can use the training size, 2K, 4K, or a custom long side while preserving the trained image aspect ratio.
- Standard SH0 3DGS PLY is available for the two Gaussian algorithms. Rectangle and Brush use non-Gaussian kernels and therefore export PNG only.
PLY results are learned from one image and prioritize its front view. Virtual Camera Sampling supports bounded tilt, but does not replace true multi-view geometry or view-dependent color.
Open index.html directly or serve the repository as static files. Before
publishing, run:
node verify-release.mjsGitHub Pages publishes the reviewed static app. Training uses the custom WebGPU
implementation; Tilt uses a pinned self-hosted
PlayCanvas Engine build. See
Third-Party Notices. This project is developed and
validated with AI assistance.
Image2SplatPaint was inspired by Image-GS and Soft Anisotropic Diagrams, while its implementation and paint-oriented design are developed independently.
- Improve faithful image representation for the Planar Gaussian and Virtual Camera paths.
- Develop Rectangle and Brush splats as distinct stylization and image-making media.
- Improve training methods for paint-oriented effects, stroke structure, and controllable visual character.
- Improve compatibility with conventional 3D Gaussian Splatting workflows.
Image2SplatPaint is released under the MIT License.