Animated glyphs¶

pl.blender.export_animation_blend(..., bake_glyphs=True) extends the GN-instanced glyph path with per-frame animation. The bridge samples each registered GlyphSpec's source dataset at every frame of the timeline, packs the per-channel values into float-precision images embedded in the .blend, and injects a small Geometry Nodes sub-graph upstream of the existing instancer so the saved file plays back natively — no Python script, no auto-execution prompt.

A moving vector field¶

import numpy as np
import pyvista as pv

# Static grid of points; the per-frame updater rotates the orient vectors.
xs = np.linspace(-1, 1, 6)
xx, yy, zz = np.meshgrid(xs, xs, [0.0], indexing="ij")
points = np.column_stack([xx.ravel(), yy.ravel(), zz.ravel()]).astype(np.float32)

cloud = pv.PolyData(points)
cloud["vec"] = np.tile([1.0, 0.0, 0.0], (cloud.n_points, 1)).astype(np.float32)
cloud["mag"] = np.ones(cloud.n_points, dtype=np.float32)

pl = pv.Plotter(off_screen=True, window_size=(960, 720))
pl.blender.add_glyph(
    cloud, geom=pv.Arrow(), orient="vec", scale="mag", factor=0.3
)
pl.camera_position = [(3.5, -3.5, 3.5), (0.0, 0.0, 0.0), (0.0, 0.0, 1.0)]

def update(frame: int) -> None:
    # Spin the orient vectors around +Z and pulse the scalar mag.
    theta = frame * (2.0 * np.pi / 24)
    cloud["vec"] = np.tile(
        [np.cos(theta), np.sin(theta), 0.0], (cloud.n_points, 1)
    ).astype(np.float32)
    cloud["mag"] = (0.5 + 0.5 * np.sin(theta)) * np.ones(
        cloud.n_points, dtype=np.float32
    )

pl.blender.export_animation_blend(
    "vector_field.blend",
    update,
    frames=range(24),
    fps=24,
    bake_camera=False,
    bake_glyphs=True,
)

Open vector_field.blend and press Space — the arrows rotate and pulse without any Python evaluation. The animation rides on keyframed image samples in the GN modifier.

What gets baked¶

Each channel is baked independently. A channel whose per-frame values are constant across the entire timeline is detected and skipped so the saved tree stays clean (no Set Position node, no image, no sampler) and the existing static value continues to drive that channel.

Sampler topology¶

Inside the GN tree, upstream of the existing instancer:

Index ─┐
       ├─ (i + 0.5) / N_points → u
       │
Scene Time ─ (frame - frame_start + 0.5) / N_frames → v
       │
Combine XYZ (u, v) → Image Texture (packed)
       │
       └─ → Set Position / Store Named Attribute
                 ↓
            Existing instancer

The Image Texture node uses Closest interpolation so each frame reads the exact pixel for the current Scene Time, not a temporal blend across rows.

Constraints and caveats¶

• Constant topology. All frames must have the same N_points. Topology-changing updaters aren't supported (no warning yet — the bake will just produce a malformed image).
• No cell-data scalars. orient and scale must live on point_data (the add_glyph API requires this for the static path too).
• Float-precision images. Each channel is a 32-bit float OpenEXR packed inside the .blend. Per-channel size scales as 4 bytes × N_frames × N_points × 4 channels. A 24-frame × 10,000- point bake is ~3.7 MB per channel.
• Scene Time vs. frame remap. Playback reads Scene Time.Frame directly, so users who shift scene.frame_start after export will see the animation scroll past the baked range; re-bake or adjust the GN frame_start subtract constant if you need to remap.