Skip to main content
Glama
mirage-gaussifier-mcp

mirage-gaussifier-mcp

Official
by Gaussifier

mirage-gaussifier-mcp

An MCP server that turns a photo into an editable image made of Gaussians, and lets an assistant edit it: swirl, magnify, recolour, duplicate, tilt parts out of the picture into 3D, and record the whole editing process as a GIF.

  • Gaussifier places the initial Gaussians in a fraction of a second.

  • MiraGe (ICML 2025) fits flat 3D Gaussians on the image plane. Each one is controlled by a triangle, so any edit is a deformation of those triangles and the picture follows.

editing a lighthouse photo through the MCP server

A quiet afternoon at the lighthouse gets a sunset, blue roofs and a whirlpool; then the lighthouse steps forward, gets a free twin, and grows a big glowing head to stay the taller one. Every step is one MCP tool call; how it was edited lists them, and the plan replays it.

How it works

  1. Gaussify. Gaussifier's equal-mass placement of the image becomes MiraGe's initial point cloud on the image plane, and MiraGe fits it with a mirror camera (about 15 s for 3000 iterations of a 768×512 image on an RTX 5090). Starting from Gaussifier gives a compact model: on Kodak 23 at 3000 iterations it ends with 81k Gaussians (40.9 dB), against 172k (47.2 dB) from MiraGe's default random start.

  2. Edit. Each Gaussian is a triangle (its centre and two scaled axes). Selections pick Gaussians by region or colour; edits move the triangle vertices (or change colour and opacity), and the renderer rebuilds every Gaussian from its triangle. Edits stack, can be undone, and each one takes a strength from 0 to 1, so it can be animated.

  3. Record. While recording, every edit is captured as a short eased animation with a caption. The GIF opens with the Gaussians growing from dots into the photo and can end with a 3D turn of the result.

Related MCP server: Blender Antigravity MCP

Installation

Requirements: Linux, an NVIDIA GPU, CUDA 12.8 or newer with nvcc, CMake, git, curl and uv.

git clone <this repository> mirage-gaussifier-mcp
cd mirage-gaussifier-mcp
bash scripts/setup.sh

scripts/setup.sh fetches MiraGe (pinned commit, with two small compatibility fixes for current compilers and Pillow) and Gaussifier v0.12.0 with its released weights into third_party/, builds the CUDA extensions and Gaussifier's runtime for the local GPU, and creates the environment.

Use it from an MCP client

The server speaks MCP over stdio:

{
  "mcpServers": {
    "mirage-gaussifier": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/mirage-gaussifier-mcp", "mirage-gaussifier-mcp"]
    }
  }
}

Put this in your client's MCP configuration (for example .mcp.json in a project, or the desktop app's settings). Then ask, for example: "Gaussify kodim23.png, make the left parrot green, swirl the right one, and record it as a GIF."

To script it without a client, scripts/mcp_session.py runs a JSON list of tool calls in one session and saves every preview:

uv run python scripts/mcp_session.py docs/demo/<name>_plan.json --out previews/

Tools

Tool

What it does

gaussify_image

Image → editable Gaussian image (init="gaussifier" or "random", iterations, max_side). Reuses an existing fit unless refit.

render

Current result, optionally with the image plane turned in 3D (yaw, pitch).

select_region

Name a set of Gaussians: box, ellipse, polygon (fractions of the image), color (rgb, tolerance), mask_image, all; combine with add / subtract / intersect. Returns the count and a highlighted preview.

edit

Apply one edit to a selection (below).

list_operations

The edits and their parameters.

undo, reset, history

Edit stack.

start_recording, stop_recording

Record the editing process as a GIF (intro, captions, 3D finale).

animate_edit, orbit

Loop one edit, or turn the result in 3D, as a GIF.

export_pseudomesh, import_pseudomesh

The triangles as an OBJ, to deform in Blender and bring back (MiraGe's workflow).

Edits: translate, rotate, scale, flip, bulge (magnify or pinch), swirl, wave, tilt3d (turn a part out of the image plane), lift (toward the viewer), recolor (hue, saturation, brightness, tint), fade, duplicate. Positions and sizes are fractions of the image.

Limits: moving or shrinking a region leaves an empty (black) area behind it, since nothing is inpainted; region selections include the background inside them, so intersect a region with a colour selection (or use a polygon) to isolate an object.

Licences

This repository's code is MIT licensed. MiraGe, which scripts/setup.sh downloads and this server imports, is distributed under its authors' Gaussian-Splatting licence for non-commercial research and evaluation use only (see third_party/MiraGe/LICENSE.md), and so is any use of this server with it. Gaussifier is used through its public release.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables AI assistants to drive a layered raster image editor with painting, shapes, text, filters, and internet-sourced images and fonts via MCP tool calls, with token-efficient rendering and export.
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables creating layered 3D reliefs from selected parts of a photo, producing GLB files and a self-contained HTML viewer that tilts with the pointer, with interactive approval of the segmentation before writing files.
      8 npm
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to model, sculpt, rig, animate and bake particle effects by editing a semantic scene document, with built-in software rendering, measurable critics that flag floating or broken geometry, and game-ready glTF export. It exposes templates, atomic edit batches, motion and effect previews, undo history, and no GPU, browser, or Blender required.
      10 npm
      Apache 2.0