threedim: new 3D asset, scene-graph and object processes (from #2109 stack) by jcelerier · Pull Request #2122 · ossia/score

jcelerier · 2026-07-14T19:51:49Z

The first cluster of new 3D objects from split/threedim, on top of the interop stack (#2121). These processes are cross-coupled (they share loader/scene headers, so intermediate subsets don't build) and therefore land as one coherent unit — the same way the scene rework did.

Stacked on #2121; retargets to master as the stack merges.

What it adds

Asset & geometry loaders: glTF/FBX import (via new fastgltf and ufbx submodules), unified AssetLoader, image/buffer/texture helpers, ObjLoader→GeometryLoader rename.
Cameras, lights, transforms: Camera, CameraArray, CameraSwitch, Light, Transform3D.
Scene-graph family: scene preprocessor, filters, switch/selector/group, duplicator, from-meshes, inspector, resource routing.
Buffer/texture inject & extract into/out of scene resources.
Materials & rendering: PBRMesh, MaterialOverride, ConfigurePrimitive, Instancer, shadow-cascade setup.
Animation: AnimationPlayer, humanoid retargeting, inverse kinematics.
Text: text-to-mesh and text-to-texture.

The existing ModelDisplay, RenderPipeline and Splat nodes keep their current scene-aware versions (the old Splat/ node is intentionally preserved). The point-cloud PrimitiveCloud family lands in a follow-up PR (it needs the spz/zstd submodules).

Validation

Builds cleanly against the stack (all plugins).
ctest: 17/17.
Process + device factory sweeps under ASAN + UBSan — the process sweep now constructs and round-trips all the new 3D process models (Camera/Light/Transform/AssetLoader/Scene*/Instancer/PBR/Animation/IK/Text/Inject/Extract): 0 memory errors, 0 undefined-behavior in score's own code, 0 crashes.

🤖 Generated with Claude Code

FIX-GLTF (GltfParser.cpp): validate every accessor's byte range against its bufferView/buffer BEFORE extract_primitive dereferences it — a crafted glTF with an out-of-range accessor offset/count read past the buffer (heap OOB). Adds accessor_within_bounds / range_within_view / buffer_source_byte_size guards + an all-accessors sweep in validate(). Test: test_regression_splat_reload (Splat ports stable across save/reload). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014rZgzE8JjWvHDtaVUhxpLE

test_unit_image_loader (22 cases / 881 assertions): score::gfx decodeImageFrom{Path,Memory} (gfx TextureLoader) + Threedim LDR ImageLoader + ArrayToTexture, across PNG8/16, JPEG, BMP, TIFF8/16, WebP, GIF, TGA, PPM/PGM, SVG, ICO — exact dimensions, exact/toleranced pixels, RGBA byte order, straight alpha, top-left origin, 16-bit precision, and graceful rejection of truncated / garbage / empty / wrong-MIME inputs (ASAN-clean). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014rZgzE8JjWvHDtaVUhxpLE

(cherry picked from commit a012d1cbdc702366d16093fda112fb3bba4a78dd)

fastgltf's configure step downloads the simdjson single-header when no simdjson::simdjson target or system package exists; network-isolated builds (Flatpak, Nix) fail there. Vendor the pinned single-header copy (v3.12.3, what fastgltf would download) so the build is self-contained. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014rZgzE8JjWvHDtaVUhxpLE

Adds the new 3D processes that build on the scene pipeline: geometry and asset loaders (glTF/FBX via fastgltf/ufbx, image and buffer helpers), cameras, lights and transforms, the scene-graph family (preprocessor, filters, switch/selector/group, duplicator, from-meshes, inspector), buffer/texture inject and extract, PBR materials and instancing, animation with humanoid retargeting and inverse kinematics, and text-to-mesh/texture. These are cross-coupled (they share loader/scene headers), so they land as one coherent unit. The existing ModelDisplay, RenderPipeline and Splat nodes keep their current scene-aware versions. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014rZgzE8JjWvHDtaVUhxpLE

Brings the new 3D objects up to their final state, folding in the bug fixes made after they were first added: Instancer GPU use-after-free and instance-count bounds, OBJ and glTF index bounds-checks, GPU-resource release on partial-init failure, Camera stale transform-slot, extraction buffer sizing and SRB layout, cubemap face sizing and UBO slots, the spherical and triplanar projection shader fixes, animation speed and skinned-mesh deformation, and the review-round corrections. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014rZgzE8JjWvHDtaVUhxpLE

ObjectGallery --list [filter] : print every registered process (name, uuid), optionally filtered by name. ObjectGallery --filter <substr> : open a window, instantiate each matching [--seconds N] object, route its texture output to the window and play, so all the texture- producing objects can be watched at once. Built on the MinimalGUIApplication + window-device path the gfx testers use, so it drives the real engine. A handy way to try the whole object catalogue. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014rZgzE8JjWvHDtaVUhxpLE

GltfParser.cpp: validate every accessor's byte range against its bufferView/buffer BEFORE extract_primitive dereferences it — a crafted glTF with an out-of-range accessor offset/count read past the buffer (heap OOB). Adds accessor_within_bounds / range_within_view / buffer_source_byte_size guards + an all-accessors sweep in validate(). Test: test_regression_splat_reload (Splat ports stable across save/reload). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014rZgzE8JjWvHDtaVUhxpLE

test_unit_image_loader (22 cases / 881 assertions): score::gfx decodeImageFrom{Path,Memory} (gfx TextureLoader) + Threedim LDR ImageLoader + ArrayToTexture, across PNG8/16, JPEG, BMP, TIFF8/16, WebP, GIF, TGA, PPM/PGM, SVG, ICO — exact dimensions, exact/toleranced pixels, RGBA byte order, straight alpha, top-left origin, 16-bit precision, and graceful rejection of truncated / garbage / empty / wrong-MIME inputs (ASAN-clean). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014rZgzE8JjWvHDtaVUhxpLE

Deduplicates the raw QImage handling and picks up decodeImageFromMemory's RGBA8888 canonicalization and footer-less TGA fallback. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014rZgzE8JjWvHDtaVUhxpLE

On glibc < 2.40 there is no _Float16, so the R16F/RGBA16F case compiled to nothing and the texture was uploaded with uninitialized bytes. Convert with a software IEEE 754 binary16 (round-to-nearest-even) fallback instead. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014rZgzE8JjWvHDtaVUhxpLE

12 Catch2 suites pinning the fixes made while building out the 3D processes: animation speed and skinned-mesh deformation, camera state release, cubemap face clamping, compute-extraction SRB layout, extraction index and buffer lifetime, HDR extension handling, instancer clamping and stale buffers, OBJ parse crashes and shader string generation. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014rZgzE8JjWvHDtaVUhxpLE

Fulldome was one hardcoded mapping. A dome is not one projection: which r = f(theta) a venue wants depends on what the content is for, and the four classical fisheye laws are not interchangeable. equidistant r = theta / (fov/2) uniform angular resolution equisolid r = sin(theta/2) / sin(fov/4) equal-area, what a fisheye lens does stereographic r = tan(theta/2) / tan(fov/4) conformal, the "little planet" look orthographic r = sin(theta) / sin(fov/2) parallel-projection sphere Each denominator is f evaluated at the half-FOV, so theta = fov/2 lands on r = 1 and the image circle meets the frame wherever the user put the FOV. The two half-angle families reduce to fov/4 for that reason and not by accident -- it is the same normalisation, not a different one. Separate shader snippets rather than a uniform the shader branches on: the projection is fixed for a whole frame, so branching would cost every vertex what selecting a program costs once. Equidistant is unchanged from the previous behaviour bit-for-bit -- its half_fov_rad is the old proj_ratio spelled differently -- so existing domes keep the picture they had. Orthographic is only meaningful to 180 degrees: sin(theta) folds past the pole, so beyond that the rear hemisphere mirrors onto the front. Left to the user like every other framing decision here, and said so in the comment. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014rZgzE8JjWvHDtaVUhxpLE

The shaders were reachable only by writing an integer to the Camera inlet from OSC or an automation, which is not a way anyone discovers a projection. All four are baked for every draw regardless, so they cost nothing extra to list. Indices follow the projections[] order in the renderer. 1 stays equidistant so documents written when it was the only dome mode keep the picture they had. "1-pass" stopped telling the user anything once the other three arrived -- they are all single-pass -- so the entry now names the projection instead. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014rZgzE8JjWvHDtaVUhxpLE

jcelerier mentioned this pull request Jul 15, 2026

threedim: PrimitiveCloud point-cloud / Gaussian-splat family (from #2109 stack) #2123

Open

jcelerier force-pushed the planf/objects branch from 9e123ef to eb14ad5 Compare July 17, 2026 03:29

jcelerier force-pushed the plane/interop branch from 801a19d to 32ad555 Compare July 17, 2026 03:29

jcelerier force-pushed the planf/objects branch from eb14ad5 to 899f99b Compare July 17, 2026 11:58

jcelerier force-pushed the plane/interop branch 2 times, most recently from 07a445e to bc681f8 Compare July 17, 2026 14:10

jcelerier force-pushed the planf/objects branch from 899f99b to ef0e957 Compare July 17, 2026 14:10

jcelerier force-pushed the planf/objects branch from ef0e957 to 3ba4662 Compare July 17, 2026 14:29

jcelerier force-pushed the plane/interop branch from 37c89a6 to 4498ad7 Compare July 17, 2026 18:12

jcelerier force-pushed the planf/objects branch 2 times, most recently from 7c5c25d to f79c9cc Compare July 18, 2026 13:37

jcelerier force-pushed the plane/interop branch from 4498ad7 to 71731df Compare July 18, 2026 13:37

jcelerier force-pushed the planf/objects branch from f79c9cc to 3d06f50 Compare July 18, 2026 14:17

jcelerier force-pushed the plane/interop branch from 71731df to 180096e Compare July 18, 2026 14:17

jcelerier force-pushed the planf/objects branch 2 times, most recently from 3fe10b1 to 5f1b3e3 Compare July 18, 2026 20:38

jcelerier force-pushed the plane/interop branch from 180096e to ea4cf98 Compare July 19, 2026 03:14

jcelerier force-pushed the planf/objects branch 2 times, most recently from 582ebb1 to df39bc1 Compare July 19, 2026 21:59

jcelerier force-pushed the plane/interop branch 2 times, most recently from 76333c0 to ca08bb3 Compare July 20, 2026 23:09

jcelerier force-pushed the planf/objects branch from df39bc1 to c59c898 Compare July 20, 2026 23:09

jcelerier force-pushed the plane/interop branch from ca08bb3 to b009f55 Compare July 21, 2026 04:55

jcelerier force-pushed the planf/objects branch 2 times, most recently from d81056a to d8d137a Compare July 21, 2026 06:13

jcelerier force-pushed the plane/interop branch from b009f55 to 63f7669 Compare July 21, 2026 06:13

jcelerier force-pushed the plane/interop branch from 6a69de3 to cfe9c11 Compare July 31, 2026 00:17

jcelerier force-pushed the planf/objects branch from 15d5c57 to 30d7f59 Compare July 31, 2026 00:17

jcelerier force-pushed the plane/interop branch from cfe9c11 to b68d334 Compare August 11, 2026 23:48

jcelerier force-pushed the planf/objects branch from 30d7f59 to 3049c13 Compare August 11, 2026 23:48

jcelerier force-pushed the plane/interop branch 2 times, most recently from 6b18a07 to 0a07aa0 Compare August 15, 2026 13:17

jcelerier force-pushed the planf/objects branch 4 times, most recently from 29bae02 to ddb1944 Compare August 15, 2026 22:20

jcelerier force-pushed the plane/interop branch 2 times, most recently from 4065f27 to 6b30420 Compare August 16, 2026 03:43

jcelerier force-pushed the planf/objects branch from ddb1944 to 5aaaa91 Compare August 16, 2026 03:43

jcelerier force-pushed the plane/interop branch from 6b30420 to a110d4a Compare August 16, 2026 13:56

jcelerier force-pushed the planf/objects branch from 5aaaa91 to 2d4ba38 Compare August 16, 2026 13:56

jcelerier force-pushed the plane/interop branch from a110d4a to 38fc230 Compare August 16, 2026 15:29

jcelerier force-pushed the planf/objects branch from 2d4ba38 to f9d4745 Compare August 16, 2026 15:29

jcelerier force-pushed the plane/interop branch from 38fc230 to 6ef9a7f Compare August 16, 2026 16:33

jcelerier and others added 12 commits August 17, 2026 14:31

3rdparty: add ufbx submodule

4f77a7c

(cherry picked from commit a012d1cbdc702366d16093fda112fb3bba4a78dd)

Navigation Menu

Uh oh!

Uh oh!

threedim: new 3D asset, scene-graph and object processes (from #2109 stack) - #2122

threedim: new 3D asset, scene-graph and object processes (from #2109 stack)#2122
jcelerier wants to merge 12 commits into
plane/interopfrom
planf/objects

jcelerier commented Jul 14, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Uh oh!

Conversation

jcelerier commented Jul 14, 2026

What it adds

Validation

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants