Changelog

All notable changes to this project are recorded here, grouped as Added, Changed, Fixed, Removed and Security. The project uses Semantic Versioning. Before 1.0.0, any release may change the API, the ABI and every data format.

0.9.0 (2026-11-02)

Changed

• Breaking: maniSkeleton_BlendPoses refuses a threshold below FLT_MIN. The threshold must be at least FLT_MIN (2^-126) and at most 1. A subnormal threshold, which 0.8.0 accepted and which could blend translations and scales to infinity or NaN, now returns mani_errorInvalid and leaves the output pose as it was. Graphs blend at MANI_BLEND_THRESHOLD (1e-5) and are not affected (#216).

Fixed

• Sampling a clip between two rotation samples in opposite hemispheres could give the identity rotation. A clip file may hold such samples, for example (1, 0, 0, 2^-80) and (-1, 0, 2^-80, 0); halfway between them, the interpolated rotation is too short for its squared length to be a normal binary32. It is now scaled by a power of two before it is normalized, so it keeps its direction. The builder keeps consecutive samples in one hemisphere, so built clips do not reach this case (#211).
• The skinning jobs gave a zero normal or tangent for a direction that had not collapsed when the palette's scale was far from 1: a uniform scale of about 2^-40 or 2^40 for normals, and 2^-70 or 2^70 for tangents. They now keep its direction. Only a zero direction, or one that is not finite, still comes out zero (#214).

Results change only for the inputs above, which gave the identity or a zero direction before. Every other pose, sample, skinned vertex, file byte and format version is the same, and no pinned hash moves. As with every release before 1.0.0, a graph instance, spring state or trainer checkpoint saved by 0.8.0 is refused with mani_errorVersion: keep assets, not saved states, across the upgrade.

0.8.0 (2026-10-31)

Changed

• Breaking: kinds, modes and choices are int32_t, and flag sets have uint32_t typedefs. All 33 kind, mode and choice typedefs, such as maniLoopPolicy, are int32_t; a negative value is refused like one past the last value. New flag-set typedefs are maniGraphSyncFlags, maniGraphClipFlags, maniRootAxes and maniMatchEntryFlags, and the fields that hold them take them, as does maniClipSource.loop, now a maniLoopPolicy. Rename MANI_MATCH_LOOP, MANI_MATCH_NO_JUMP and MANI_MATCH_MIRRORED to maniMatchEntryLoop, maniMatchEntryNoJump and maniMatchEntryMirrored. maniMorphTargets_GetAttribute returns -1 past the count, where it returned UINT32_MAX. Code that keeps a kind in a uint32_t variable or compares it with an unsigned value updates its types (#199).
• Breaking: two names follow the family's naming rules. maniLearnedTrainerReport.iteration is iterationCount, and maniRootMoveIdentity is maniMakeRootMoveIdentity (#198).
• Breaking: the CPU checks return bool, and maniHash64 takes an int64_t count. maniIsCpuBackendSupported and maniCheckCpu return bool instead of an int32_t 1 or 0, so a caller comparing the result with 1 tests it as a bool. maniHash64 hashes any byte count up to INT64_MAX; pass the count without narrowing it. Hash values are unchanged (#176).
• Breaking: a refused call or a "no" clears its outputs. A call on a builder, a glTF or FBX document or a trainer clears its sizes, counts, reports, structs and handles before any check, so a failed call never leaves an earlier call's value to be read as its answer. maniBuilder_CopyBuilt, maniBuilder_CopySkeletonOrder, maniGltf_GetMorphWeights, maniFbx_GetMorphWeights and maniLearnedTrainer_SaveCheckpoint also zero the caller's array up to its capacity; a short skeleton order is no longer filled as far as it goes. A too-small buffer still reports the size or count it needed. maniBuildGltfClip and maniBuildFbxClip write boundOut only when the clip builds. maniClip_GetRootMotionSource and maniInfluences_GetCenter clear their output when they return false, and the name getters of learned models and motion databases clear lengthOut when they return NULL. The headers say what each output holds after a failure (#165, #174).
• Breaking: a call refused as busy leaves its outputs. A builder or trainer call that its own allocator or task hook made is refused with mani_errorBusy and writes none of its outputs, since they may be the ones the call in progress is filling. maniOpenGltf and maniOpenFbx still set their handle to NULL (#173).
• Breaking: more calls back in from host code are refused as busy. maniGltf_SupplyBuffer, maniGltf_ReadSprings and maniGltf_ReadRagdoll return mani_errorBusy when called from the builder's allocator. maniLearnedTrainer_SaveCheckpoint and maniBuildLearnedModelFromTrainer return it when called from the trainer's allocator or task hooks. While a model is built from a trainer, maniLearnedTrainer_Train, maniLearnedTrainer_RestoreCheckpoint and maniDestroyLearnedTrainer on that trainer return it too. Each refusal counts one misuse (#187, #197).
• The headers say which outputs may be their input: maniBlendRootMotion's moveOut may be one of moves, maniLearnedModel_Project's stateOut may be query, and maniMorphTargets_ComputeTargetWeights' targetWeightsOut may be channelWeights (#195).

Fixed

• maniMorphTargets_ComputeTargetWeights gave wrong weights when its output was its own channel weights. It now computes them in place with the same bits as apart (#193).
• maniRetargetMap_RetargetPose gave a wrong pose when the target pose was the source pose, or shared memory with it. It now refuses any overlap of the two poses with mani_errorInvalid and writes nothing (#194).
• Destroying a builder or a trainer, or closing a glTF or FBX document, let the host's free hook call back into the object: a hook could build on a builder being destroyed, free it twice, or read a closing document's freed names. Every call from those frees is now refused with mani_errorBusy, and a closing document reads as empty (#187).
• Host code running while a model was built from a trainer could destroy or retrain the trainer under the build, which then read freed memory. The trainer is now marked busy for the whole build (#197).

Poses, samples, file bytes and format versions do not change. As with every release before 1.0.0, a graph instance, spring state or trainer checkpoint saved by 0.7.0 is refused with mani_errorVersion: keep assets, not saved states, across the upgrade.

0.7.0 (2026-10-29)

Changed

• Breaking: every call is named by the Maul family's naming rule. A call on one object is named maniObject_Verb: a getter maniObject_GetThing, a predicate maniObject_IsThing. The object is the library object the call takes first, such as a clip, a pose, a skeleton, a graph instance, a builder or a glTF document, even where the call only reads it: maniSampleClip is maniClip_Sample, maniClipDuration is maniClip_GetDuration, maniPoseIsDifference is maniPose_IsDifference and maniBlendPoses is maniSkeleton_BlendPoses. Creation, destruction and calls with no object keep a verb first: every maniOpen..., maniInit..., maniCreate... and maniBuild... keeps its name, and the importers' builds take that shape (maniGltfBuildClip is maniBuildGltfClip). The misuse counts are maniBuilder_GetMisuse and maniLearnedTrainer_GetMisuse, maniResultName is maniGetResultName, the CPU check is maniIsCpuBackendSupported, and the task hook types end in Function. Math on plain values keeps its name (maniRootMoveIdentity, maniComposeRootMotion, maniDecomposeAffine, maniHash64). Rename every call in the table below; the parameters stay the same, except for maniPose_AddLayers (#156).
• Breaking: maniPose_AddLayers takes the pose first. A pose a call changes in place is its object, so maniAddPoses(skeleton, pose, layers, count) is maniPose_AddLayers(pose, skeleton, layers, count). Swap the first two arguments; a compiler refuses or warns about the old order, since the two pointers have different types. The IK solvers already took the pose first and keep their order as maniPose_SolveTwoBone and the others (#156).
• The user guides and the API reference use the new names (#156).

Poses, samples and file bytes do not change; only the names do. As with every release before 1.0.0, a graph instance, spring state or trainer checkpoint saved by 0.6.0 is refused with mani_errorVersion: keep assets, not saved states, across the upgrade.

0.6.0 (2026-10-29)

Changed

• Breaking: a saved state is read only by the release that wrote it. A saved graph instance and a saved spring state start with the Maul family's 16-byte state prefix (magic, MANI_ABI_VERSION, format version, header size), so their headers are 32 bytes and every saved state's bytes change. maniRestoreGraphInstance, maniRestoreSpringState and maniRestoreLearnedCheckpoint refuse a state written by a release of another ABI version with mani_errorVersion: before 1.0.0 that is any other release, from 1.0.0 another minor release. A state saved by 0.5.0 or earlier is refused. Keep assets, not saved states, across an upgrade. Poses, samples and file bytes do not change (#134).
• Breaking: each save call reports the saved size. maniSaveGraphInstance, maniSaveSpringState and maniSaveLearnedCheckpoint take a sizeOut after the capacity, and a NULL buffer with capacity 0 measures. maniGraphStateSize, maniSpringSavedSize and maniLearnedCheckpointSize are removed: call the save with a NULL buffer instead (#129).
• Breaking: every call that fills an array takes its capacity and reports the count. maniMotionDatabaseFeatures, maniMotionDatabaseNormalization, maniLearnedModelNormalization, maniMorphTargetWeights and maniLearnedProject take a capacity and a countOut after the array; maniGltfMorphWeights and maniFbxMorphWeights take a capacity in place of the exact count, and a countOut; maniRagdollBodiesFromPose, maniRagdollJointTargets and maniBuildMask take a countOut after the capacity. maniLearnedOutput gains floatTrackCapacity and eventCapacity after its arrays. A NULL array with capacity 0 measures (#133).
• Breaking: name lengths are size_t. Every name getter (the skeleton's bones, a clip's tracks, the graph's slots and parameters, a motion database's channels, a learned model's tracks and events, morph channels and targets, a spring set's springs, and the glTF and FBX importers' names) reports a length through size_t*, and every name length a public struct carries is size_t, so a name read from one call passes into a source unchanged. A builder refuses a name longer than the new MANI_MAX_NAME_BYTES (2^31 - 1 bytes) with mani_errorLimit, at the reason and item it gives that name otherwise; before, a morph channel name past it was refused with mani_errorInvalid, and a graph event name past it was hashed as if empty (#135, #138).
• Breaking: the graph's retargeting maps are named as such. maniGraphMapCount, maniGraphMapIdentity, maniGraphMapSourceIdentity and maniGraphClipMap are maniGraphRetargetMapCount, maniGraphRetargetMapIdentity, maniGraphRetargetMapSourceIdentity and maniGraphClipRetargetMap; MANI_MAX_GRAPH_MAPS is MANI_MAX_GRAPH_RETARGET_MAPS; maniGraphClipSource.map is retargetMap, and maniGraphAssets.maps and mapCount are retargetMaps and retargetMapCount (#136).
• Every builder's reference names the MANI_MAX_NAME_BYTES refusal, and the getting-started guide gives the bound as an example of a fixed one (#155).

Fixed

• The motion database builder no longer writes past its names. Channels may name one track, and their names' total was kept in 32 bits: two channels naming one track of 2^31 + 1 bytes wrapped it, and copying the names wrote past the block it sized. Names past 4 GiB in all are refused with mani_errorInvalid and maniBuildReasonCount, before any name is read, and where size_t has 32 bits, sections past what one allocation holds are refused with mani_errorCapacity (#143).
• The mirror map builder sizes its working memory in 64 bits. Where size_t has 32 bits (WebAssembly), listed names totalling 4 GiB wrapped the room for their partners, which were then written past it; such a source is refused with mani_errorCapacity (#142).
• The graph and mirror map builders refuse names past their files' 32-bit offsets. Names or partners past 4 GiB in all were laid out with wrapped offsets, so the file built pointed at other bytes; they are refused with mani_errorInvalid and maniBuildReasonCount, and the mirror map's sections past size_t with mani_errorCapacity (#144, #151).
• An imported name past 4 GiB is refused, not cut short. The glTF and FBX importers cut a name's decoded length to 32 bits, handing the builder the first bytes of the real name; the whole name now reaches the builder, which refuses it with mani_errorLimit (#153).

0.5.0 (2026-10-28)

Changed

• Breaking: every file kind's layout and format version change, and an opener reads only its own version. The reserved words are gone from the files: the container header is 48 bytes, not 64, and the clip, retargeting map, influences, learned model, learned checkpoint, spring set, ragdoll profile and mirror map lose their zero words. Every kind's format version goes up by one: skeleton 2, clip 4, graph 7, retargeting map 3, skin 2, influences 2, morph targets 2, motion database 4, learned model 2, learned checkpoint 2, spring set 3, ragdoll profile 2, mirror map 3. An opener now refuses any format version but its own with mani_errorVersion, an earlier one included; it used to accept an earlier version and read it with the current layout. Build every stored file again from its source. Poses, samples and other outputs do not change; only hashes of file bytes move (#113).
• Breaking: results take the Maul family's values. mani_errorVersion is -7 (was -4) and mani_errorFormat is -8 (was -3). mani_errorState is removed: a call from an object's own host code (its allocator or task hooks) is refused with the new mani_errorBusy (-10) in its place, and still counts one misuse. Code that compares results by name only renames mani_errorState to mani_errorBusy (#115).
• Breaking: a fixed bound passed returns mani_errorLimit (-9). An argument, or a builder's or the trainer's source, past a bound the library publishes returns mani_errorLimit in place of mani_errorInvalid, with the same reason and item. This covers the aim, CCD, FABRIK and full-body IK solvers, the motion database search, and the skeleton, skin, influences, clip, morph targets, spring set, retargeting map, motion database, learned model, trainer and graph builders. A bound checked while opening a file is still mani_errorFormat, and a range of a parameter's meaning (a weight in [0, 1], a cone up to pi) is still mani_errorInvalid. Full-body IK scratch that is too small returns mani_errorCapacity, not mani_errorInvalid (#117).
• Breaking: the bound constants are named MANI_MAX_*. MANI_AIM_MAX_BONES is MANI_MAX_AIM_BONES, MANI_IK_CHAIN_MAX_BONES is MANI_MAX_IK_CHAIN_BONES, MANI_IK_MAX_ITERATIONS is MANI_MAX_IK_ITERATIONS, MANI_SPRING_MAX_RATE is MANI_MAX_SPRING_RATE, MANI_MORPH_MAX_ATTRIBUTES is MANI_MAX_MORPH_ATTRIBUTES, and the MANI_BODY_IK_MAX_*, MANI_MATCH_MAX_*, MANI_LEARNED_MAX_* and MANI_GRAPH_MAX_* names are MANI_MAX_BODY_IK_*, MANI_MAX_MATCH_*, MANI_MAX_LEARNED_* and MANI_MAX_GRAPH_*. A name that keeps the old form (MANI_AIM_MAX_CONE, MANI_SPRING_MAX_STEPS, MANI_GRAPH_LEARNED_MAX_STEPS) is a range or an internal cap, not a bound refused with mani_errorLimit (#117).
• Breaking: public structs have no reserved field. The field is removed from every public struct, and builders no longer refuse a source for a nonzero one. Remove it from initializers. maniFileResult is {result, tag, offset}: tag names the section the first wrong byte lies in, 0 in the header and table, so positional initializers take the tag second (#109).
• Breaking: a root graph's slots are checked against the root's own lists. maniBuildGraph refuses, with maniBuildReasonNode at the node's place, a root node that names a clip, sync group, database, model or spring set slot past the root's own list. In a graph with a subgraph such a node used to build with the subgraph's slot of that number: with 1 clip in the root and 1 in the subgraph, a root clip node naming slot 1 played the subgraph's clip. Only a host's own maniGraphSource structs reach this; the JSON form resolves each name within its own graph. Every other graph builds to the same bytes. To migrate, list in the root every asset and sync group its nodes name (#93).
• MANI_MAX_INFLUENCES (64 weighted influences on a vertex) and MANI_MAX_GRAPH_DEPTH (16 levels of subgraphs) are public (#117).
• On WebAssembly, the public calls no longer go through the floating-point environment's out-of-line entries, which have nothing to switch there, so the static library is about 14.7 KB smaller. Results do not change (#101).
• The user guide's getting started page explains mani_errorLimit, mani_errorBusy and the one format version a release reads, and its examples drop the reserved field and use the new constant names.

Fixed

• The FBX importer bounds a file's property count. A file with 2^32 - 1 properties or more wrapped the count, and the importer wrote past the property array it had allocated from it. Such a file is now refused at the property past the bound; every file under it reads as before (#105).
• A motion database file whose entry is too long for a frame count and stores 0xFFFFFFFF as its count is refused at that entry's count. It was still refused, but at a later, wrong place. Valid databases open and build as before (#107).

0.4.0 (2026-10-25)

Changed

• Breaking: a subgraph's lists and slots are checked like the root's. maniBuildGraph refuses with maniBuildReasonNode, at the node's place, a node inside a subgraph that names a clip, sync group, database, model or spring set slot past the subgraph's own list. Such a graph used to build when the root had a slot of that number, which the node then named: with 2 clips in the subgraph and 9 in the root, a clip node naming slot 3 played the root's clip 3. It also refuses with maniBuildReasonCount a subgraph with a list whose count is not zero and whose pointer is NULL; its clip, sync group, database, model and spring set lists used to be read through that pointer when a node named one of their slots. Only a host's own maniGraphSource structs reach either case; the JSON form resolves and checks its lists. Every other graph builds to the same bytes. To migrate, list in each subgraph every asset and sync group its nodes name (#77).
• On 32-bit x86 the library is built with SSE2 math (-msse2 -mfpmath=sse), so it needs a processor with SSE2, and every build now leaves trapping math at the compiler's default, from the family's compiler flags. 64-bit builds' tests and pinned hashes pass unchanged (#73).

Fixed

• maniBuildGraph refuses a subgraph's sync group whose name is NULL with maniBuildReasonCount at item 0, before anything is laid out, as it refuses a subgraph's list that is not given. A root's own nameless group is still refused by the check with the sync group list's Name reason. Only a host's own maniGraphSource structs reach this; the JSON form always gives a name. Every graph that built before builds to the same bytes (#81).

0.3.0 (2026-10-25)

Changed

• Breaking: a graph's time scales are bounded. On any path from the output, the product of the time scales' sizes must be at most MANI_GRAPH_MAX_TIME_FACTOR (1e6), which is far past any playback rate a graph needs. A scale that reads a parameter counts at the larger size of the parameter's bounds, and a time scale may read only a parameter with bounds. Scales on parallel branches never multiply. maniBuildGraph refuses a source past either rule with maniBuildReasonScalar, and maniOpenGraph refuses such a file with mani_errorFormat. Each node record now holds its time factor, so the graph format goes from 5 to 6, and maniOpenGraph refuses an earlier format with mani_errorVersion. To migrate, give bounds to each parameter a time scale reads, and build every graph file again. Results do not change for any graph still taken. (#66)

Fixed

• A clip node's step that is not finite moves nothing, and the marker walk's count is bounded by two loops of markers. Results do not change for steps shorter than about 2^53 clip lengths. (#64)
• A graph's binary32 timers saturate at FLT_MAX: the fade time of a machine and of a one-shot, and an inertialize node's step and elapsed time. Every timer value below FLT_MAX gives the same bits as before. (#67)

0.2.1 (2026-10-24)

Changed

• The runtime's core (sampling, blending, the model pass, skeletons and root motion) is about 1.0 KB smaller at -O2 with GCC: 68,562 bytes, against ozz-animation's 69,698. The skeleton's opener and the file checks are compiled for size under GCC and Clang. With GCC, opening a skeleton takes about 7% more instructions and opening a compressed clip about 9% more. Opening a full clip or a motion database does not change, and neither does any call made every frame. Results do not change. (#29)

Fixed

• The host's allocator and task hooks run in the host's own floating-point setting, like the rest of the host's code. In 0.2.0 they ran in the library's setting while a public call was in progress. (#27)

0.2.0 (2026-10-24)

Changed

• Breaking: motion matching searches a motion database through an exact k-d tree, in place of bounds over runs of frames. Every search returns the same frame and cost as before. On large databases a search takes far fewer instructions: about 6 times fewer at 100,000 frames and up to 16 times fewer at 300,000 frames with the default leaf. On small databases with many features (around 5,000 frames and 45 features) it can be slightly slower. The index takes about 49 bytes a frame at the default leaf, up from about 17. The motion database format version is 3, and maniOpenMotionDatabase refuses any other format version with mani_errorVersion. Build databases again from their sources, along with the graphs that name them, since a database's identity changes. A learned trainer's checkpoint does not resume against a rebuilt database. maniMotionDatabaseSource gains leafFrames (8, 16 or 32 frames a leaf, 0 for 16) and a reserved word that must be 0, so positional initializers must name them. (#24)

Fixed

• A host that runs its threads with flush-to-zero, denormals-are-zero or another rounding mode now gets the same bits as every other host. Every public call that computes switches to the library's floating-point setting and restores the host's on return, and so do the learned trainer's tasks on the host's worker threads. Results in the default setting do not change. (#26)

0.1.0 (2026-10-24)

Changed

• Compressed clips keep each kind of sub-track together and decode several sub-tracks at once: sampling at a random time takes about 28% fewer instructions and a forward cursor about 15% fewer, and poses and float tracks are the same bits as before. The clip format version is 3, and maniOpenClip refuses a clip of any other format version, compressed or full precision, with mani_errorVersion. Build stored clips again from their sources, or compress them again from full-precision clips made with this release.

Fixed

• A learned motion matching node whose model gives a value past the float range, or whose root velocity moves past it over the step, no longer passes it on: maniUpdateGraph, maniEvaluateGraph and maniEvaluateGraphSubset complete the call with valid, finite outputs and return mani_errorFormat, and maniLearnedStep, maniLearnedProject and maniLearnedDecompress return mani_errorFormat instead of success. maniGraphNodeView's reserved word is now failed.

0.0.1 (2026-10-24)

The first release: skeletal animation from import to the skinned vertex, with the same results on every platform, compiler and SIMD backend, and no allocation at run time.

Added

• Skeletons, poses in local and model space composed as glTF composes nodes, and the polar decomposition of model matrices.
• Clips on a uniform grid sampled at the same cost at any time, with float and discrete tracks, events and sync markers; a forward cursor with the same results, in a faster form and a compact one.
• Clip compression within an error measured on a shell around each bone, close to ACL's sizes at a third of its error.
• Blending with per-bone masks, pose differences and additive layers and clips, and root motion taken out at build and read as moves across any number of loops, forward or backward.
• Animation graphs: clip, blend, mask, additive, select, time scale and seek nodes; 1D, triangulated and banded blend spaces; state machines with conditions on parameters, time, a clip's phase, end or events; one-shots; inertialization and dead blending; sync groups by phase or markers; subgraphs; a JSON form; saved and restored instances; node inspection for debuggers; evaluation of a subset of bones.
• Motion matching: motion databases with weighted feature channels, an exact pruned search, a trajectory predictor and a graph node; learned motion matching with a model format, a graph node and an in-library trainer whose results are the same on every platform and with any number of threads.
• Inverse kinematics: two-bone, aim and look-at over a chain, CCD and FABRIK chains with cone, twist and hinge limits, full-body XPBD and SDLS solvers, a turn limit node, each also a graph node.
• Spring bones as VRMC_springBone 1.0 defines them, with its extended colliders and limits, the same motion at any frame rate.
• Ragdoll profiles mapping bones to a physics engine's bodies and joints, bodies and motor targets from poses, poses from bodies, and get-up placement; a pose input for graphs.
• Retargeting between skeletons with different names, chain lengths, rests and proportions, and mirroring on rigs whose sides do not have mirrored axes, for poses, baked clips, graphs and motion databases; JSON forms for both maps.
• Skinning: skins, palettes, linear, dual quaternion and blended skinning and centers of rotation over one influence format, as jobs over strided vertex streams; morph targets with in-betweens.
• glTF 2.0 import (JSON and GLB, EXT_meshopt_compression, KHR_meshopt_compression, KHR_draco_mesh_compression, KHR_animation_pointer, VRMC_springBone and the OMI_physics_* extensions) and FBX import (binary and ASCII, versions 3000 to 7700): skeletons, skins, influences, morph targets and clips.
• Validation of every file, JSON text and saved state as it is opened, in every build: bad data is refused with a status and the offset of the first wrong byte, and counts no misuse on the builder; a misuse count reports only calls a correct program never makes.
• A CPU check for the AVX2 build, a portable scalar build with the same results, a CMake package and a pkg-config module.
• Support for Linux on x86-64 and arm64, macOS, Windows with clang-cl, and WebAssembly.