Manim Vision adds spatial collision awareness to Manim Community scenes. It watches your scene as add, play, and remove run, detects meaningful overlaps, and emits collision timelines that an LLM can use to fix layout problems without reading video frames.

• Hooks a live Scene without changing your source class at import time.
• Converts tracked VMobject geometry to Shapely shapes and checks overlap with an STRtree.
• Suppresses obvious noise such as glyph-on-glyph kerning, centered text-in-cell layouts, and tiny dust overlaps.
• Tracks overlap intervals, so a collision that persists for 100 plays is reported once with a start time and end time.
• Labels collisions with scene-meaningful names instead of raw memory ids.
• Aggregates leaf-precise hits into scene-level collision episodes, so one visual problem becomes one readable record.
• Produces a compact scene summary by default, with optional human-readable or silent modes.
• Can replay any monitored scene as a separate collision-overlay video for visual debugging.

pip install manim-vision

Dependencies are declared in pyproject.toml and include manim, shapely, wrapt, numpy, and jsonschema.

Call ManimVision.monitor(self) early in construct(), then shut it down at the end so queued collision work can flush.

from manim import BLUE, RED, Create, Circle, RIGHT, Scene, Square
from manim_vision import ManimVision


class MyScene(Scene):
    def construct(self):
        ManimVision.monitor(self)  # default: compact LLM summary

        circle = Circle(radius=1, color=BLUE)
        square = Square(side_length=1.2, color=RED)
        square.next_to(circle, RIGHT, buff=0)

        self.add(circle, square)
        self.play(Create(circle), Create(square))

        ManimVision.shutdown(self)
        print(ManimVision.results(self))

ManimVision.monitor(scene, output_mode=...) accepts one optional mode:

• "llm": default. Collects collision events during the scene and writes one semantic v2 summary at shutdown to media/manim_vision/<SceneName>_check_digest.jsonl.
• "human": collects the same events but writes a readable interval report for developers.
• "silent": writes nothing to disk or stdout and keeps results available through ManimVision.results(scene).

Example:

ManimVision.monitor(self, output_mode="silent")
# ...
ManimVision.shutdown(self)
summary = ManimVision.results(self)

To generate a diagnostic overlay video after a run:

overlay_path = ManimVision.render_overlay(
    r"C:\path\to\scene.py",
    "MyScene",
)
print(overlay_path)

By default, Manim Vision writes under your Manim media directory:

• media/manim_vision/<SceneName>_check_digest.jsonl: semantic collision timeline with object names, intervals, contact spans, hotspots, and fix hints.
• media/manim_vision/<SceneName>_finalcontextcollisionreport.json: compact LLM-facing report for feedback loops.
• media/manim_vision/<SceneName>_overlaydata.json: geometry-rich replay payload used to render diagnostic overlay videos.
• media/manim_vision/<SceneName>_spatial_log.txt: human-readable collision timeline in human mode.
• media/manim_vision/<SceneName>_spatial.jsonl: legacy per-event JSONL, only when MANIM_VISION_PER_PAIR_JSONL=1.

Override the directory with MANIM_VISION_REPORT_DIR.

• ManimVision.monitor(scene, output_mode="llm")
• ManimVision.shutdown(scene)
• ManimVision.results(scene)
• ManimVision.render_overlay(scene_or_script, scene_name=None, ...)

Exceptions re-exported from manim_vision:

• ManimVisionError
• ManimVisionGeometryError
• ManimVisionSchemaError
• ManimVisionProxyError

1. ManimVision.monitor(scene) attaches an engine, solver, dispatcher, lock, and worker executor to the live scene instance.
2. A mixin is inserted at runtime so add, play, and remove can register, resync, and deregister geometry.
3. Collision checks run on a single-worker background executor under a scene lock.
4. Raw leaf collisions are filtered, merged by semantic owner pair into one continuous overlap episode, and flushed on shutdown.
5. The default v2 report keeps exact timings and contact locations while hiding animation-internal path noise from the final LLM-facing output.

• Analysis is 2D and based on overlap area, not full 3D physics.
• Touching edges without positive overlap are ignored.
• Some concave cases use centroid-based fallbacks for fix hints.
• Internal lock-bearing state stays on proxies, so Manim creation-style animations still deep-copy safely.

python -m venv .venv
.venv\Scripts\activate
pip install -e ".[dev]"
python -m pytest

MIT. See LICENSE.