Animation and performance

2D/3D animation, rigs, retargeting, motion capture, facial animation and lip sync

• articleWhat does a Houdini Vellum cloth simulation carry into an Alembic file for rendering, and which attribute and packing details can change on export?
  Houdini's Alembic ROP writes only the extra attributes whose names match its point/vertex/primitive/detail patterns and turns primitive groups into face sets, which Alembic supports only on polygon and subdivision meshes; Vellum Unpack depends on the vellum_type attribute and does not update id attributes properly, so ids and constraint context should be checked before baking cloth to Alembic, which stores results rather than the simulation.
• articleHow do these tools implement animating 'on twos' (and threes/fours), and what changes in frame count and project frame rate for each method?
  In OpenToonz, Step 2/3/4 re-exposes selected drawings to create twos/threes/fours (adding frames), while Each 2/3/4 thins a selection to every second/third/fourth drawing (removing frames); because tools also offer capture-time holds and playback-rate changes, 'on twos' only has a fixed meaning when the project frame rate and whether frame count changes are stated.
• articleWhat does the 2024 ICAT-EGVE poster map between FACS action units and ARKit blendshapes, in which direction, and what does it measure?
  The 2024 EGVE poster learns ARKit's 52 blendshape weights from 17 OpenFace action-unit intensities with a GRU, trained on 3,780 videos of nine Japanese speakers reading emotion-related sentences, and reports R2 about 0.75 with activation precision about 0.90 and recall about 0.85; it maps AUs to blendshapes for speech, so it is not a general AU-name lookup table and its population and task are narrow.
• articleHow does the C3D file format store 3D points and analog channels, and how must scale, rate and units be handled?
  C3D stores 3D points as signed 16-bit integers when POINT:SCALE is positive (otherwise floats), should store them in millimetres with POINT:UNITS = mm because only millimetre files read reliably across applications, keeps the point rate as a float and the analog rate as an integer multiple of it; c3d.org describes C3D as a public-domain format. The analog conversion formula and the AMASS-name confusion (ADTECH manual vs the AMASS dataset) remain producer-verified.
• articleWhat terms and limits do the MediaPipe face model cards state?
  Google's MediaPipe face model cards (BlazeFace short range, Face Mesh V2, Blendshape V2) each declare the model licensed under Apache 2.0, name AR entertainment from front-facing monocular video as the primary intended use, and list out-of-scope uses including human life-critical decisions; Blendshape V2 outputs 52 coefficients in [0,1]. The cards are the licence statement for these models; the bundled .task file was not inspected.
• articleWhat can Alembic ICurves carry for hair and groom data, and what do Blender's Alembic and USD hair exports each drop or change?
  Blender exports the same particle hair as linear, non-periodic, no-basis curves with UVs and normals (optionally child strands) to Alembic, documented as zero-width curves, but as cubic bspline or catmullRom pinned BasisCurves without widths to USD. Alembic ICurves itself can carry widths, UVs, normals, velocities, orders and knots, so the narrower content is Blender's export choice, not a format limit.
• articleHow does the USD BasisCurves schema at v26.08 describe hair-style curves, and what does Blender 5.2.2 write when it exports particle hair and Curves objects to USD?
  Blender v5.2.2 writes particle hair to USD as BasisCurves (cubic bspline, or cubic catmullRom with pinned wrap) with points, curve vertex counts and a display colour, and no widths. The USD schema description of BasisCurves for hair, the Curves-object writer (widths from radius, UV primvars, velocities) and the parents-only manual note remain producer-verified.
• articleWhen loading FBX with ufbx, which space-conversion mode should be used to normalise axes and units, and what pitfalls does the header document?
  ufbx converts FBX coordinates and units only when target_axes or target_unit_meters is set, using one of three modes: put the conversion on the root node, compensate per node with adjust pre-rotation/pre-scale, or additionally scale geometry. The scene settings still report the file's original axes, and FBX cameras are saved +X forward and +Y up unless target_camera_axes overrides it. Which mode to use depends on whether downstream code rebuilds transforms from raw properties (then adjust transforms must be accounted for).
• articleWhich axis, handedness, unit and time conventions does each of glTF, OpenUSD, FBX (via ufbx), Blender, Unreal, Unity, Maya and Houdini document, and where do they have to be converted?
  Default conventions differ: glTF is right-handed, +Y up, asset front facing +Z, in metres; OpenUSD is right-handed with a per-stage upAxis (Y or Z; files should always set it) and a metersPerUnit scale; FBX (via ufbx) usually uses centimetres and its 'front' axis is the opposite of forward; Unreal's editor is left-handed Z-up; Blender works Y-forward Z-up and its USD exporter can convert orientation. Any conversion must reconcile handedness, up/forward axis and unit scale explicitly.
• articleWhat do the OpenXR registry and the META viseme chapter not define about visemes (ratification, timing, mapping to blend shapes)?
  The viseme extension is a vendor (META) extension without a ratified attribute in the 1.1.65 registry, and its chapter defines no viseme-to-blend-shape or ARKit mapping, no output rate and no latency; an agent mapping these visemes to a rig must supply its own mapping and timing assumptions.
• articleWhat does XR_META_face_tracking_visemes define (viseme list, value semantics, validity, dependencies) in the OpenXR registry and spec source?
  OpenXR (release 1.1.65) defines the Meta viseme extension XR_META_face_tracking_visemes (number 542, layered on XR_FB_face_tracking2) with 15 visemes (SIL, PP, FF, TH, DD, KK, CH, SS, NN, RR, AA, E, IH, OH, OU) reported as probabilities that sum to 1 when valid; it is meant to match common 15-viseme sets and the older Oculus Lipsync visemes, and apps can request visemes without blend shapes by setting weight and confidence counts to zero.
• articleWhere can a digitised blank paper animation exposure sheet (X-sheet) from a named studio and date be found in a public collection, and what columns does it show?
  The only dated, named-studio X-sheet found in public catalogues is an Internet Archive user upload (UPA New York, 1952; one JPEG, no licence, no text layer, provenance a discontinued blog), so whether it is blank and what columns it has are unestablished without viewing the image. A generic classic layout (Action, Dial, five cel-level columns, Background, Camera; 96 rows = 4 seconds) comes from a 2023 CC BY-NC-ND 4.0 teaching item with no studio or date. Institutional finding-aid examples were not found in limited queries (not proof of absence).
• articleWhat do directors, actors and a cinematographer say first-hand about how performance changes with shot size, lens and coverage, with concrete examples?
  Directors and actors describe close-ups as performance tools, not a size rule: Demme shot super-tight close-ups looking over and into the lens with tight over-the-shoulders to move in and out of a character's point of view; Attias wants each shot specific rather than enlarging one wide frame, including a deliberate wide-angle close-up; Glenn Close says a well-made close-up keeps the audience emotionally involved and that silence can carry a scene on film.
• articleAcross the capture-to-render chain, which time base, unit and naming assumptions must be written down because no format in the chain carries them for the next step?
  Across motion and render handoffs, time, units/axes and names are carried differently by each format (ARKit metres; Live Link Face timecode in recordings; Alembic sample times; the USD Alembic reader assuming Y-up and leaving units unset; UsdRender and EXR names as site conventions), so each step needs a user-authored mapping; this is the producer's synthesis over verified statements.
• articleWhat rotation order, scale, rest pose and axis conventions does Blender's BVH exporter use, and how does the importer differ?
  Blender's BVH exporter writes rest-pose offsets relative to the parent scaled by global_scale, 6 channels (position then rotation) for root or disconnected bones and 3 rotation channels for connected bones, Euler angles in degrees in each bone's rotation order, and Frame Time from the scene fps; unlike the importer it offers no forward/up axis conversion, so receiving tools must handle axis differences themselves.
• articleWhen animation leaves the DCC as UsdSkel or an Alembic cache, which of joint order, bind/rest pose, topology, velocity, time base and units are stated by the specs and what does the Alembic-to-USD reader do with the rest?
  UsdSkel stores joints in an ordered list with world-space bind and local-space rest transforms (rest may be omitted only with a complete, non-sparse animation; animation joint order may differ from the skeleton's). Alembic marks topology changes by full samples and stores explicit time samples; OpenUSD's Alembic reader takes the frame rate from archive metadata, assumes Y-up when none is authored and never sets metersPerUnit, so units and up-axis need checking after conversion.
• articleWhat does Learned Motion Matching (Holden et al. 2020) replace, what data and training did it use, and what limits do the authors report?
  Learned Motion Matching (2020) replaces the stored animation database and nearest-neighbour search of motion matching with three small networks (Decompressor, Stepper, Projector) so the runtime keeps no pose database; memory is the stated motive. Training runs on CPU for up to 500,000 iterations (a few hours for results, overnight for best). Reported scenarios span 22,398 to 694,272 frames, but the 694k Bear set was augmented from about one minute (3,581 frames) of key-framed animation. In the user study subjects struggled to identify the reference system, with a slight responsiveness preference for basic motion matching and no quality preference; the authors found Projector accuracy had the largest impact on quality.
• articleAt UniVRM v0.131.4, how does the VRM 0.x to 1.0 migration code map each meta field and what defaults does it set?
  UniVRM's VRM 0.x to 1.0 meta migration begins from the most restrictive 1.0 licence values and only loosens what 0.x fields state; the 0.x licenseName (e.g. CC0, CC BY) is not mapped to any 1.0 field (TODO), so 1.0 fields with no 0.x source (redistribution, modification, credit, political/religious use, etc.) stay restrictive after migration, and unrecognised values throw.
• articleHow do the open-source Maya Alembic exporter, maya-usd and Houdini handle sub-frames and non-integer frame rates like 23.976 and 29.97?
  maya-usd v0.37.0 maps Maya's 23.976, 29.97 and 29.97 drop-frame units to exact rational rates (24000/1001, 30000/1001) and writes the value as both timeCodesPerSecond and framesPerSecond, unlike Blender's USD exporter which writes only timeCodesPerSecond.
• articleWhen Blender exports USD or Alembic from a 23.976 fps scene, which time values and sub-frame samples does it actually write?
  Blender v5.2.2's USD exporter writes stage timeCodesPerSecond from the scene rate (e.g. 24000/1001) and exports whole frames only, so it cannot emit sub-frame motion samples, and it does not write framesPerSecond; its Alembic exporter writes shutter-based sub-frame samples and stores the scene fps as FramesPerTimeUnit metadata.
• articleWhat do Epic's public docs say about motion matching database size, search mode and indexing, and where are the GDC 2024 numbers?
  Epic's UE 5.8 Motion Matching page gives only qualitative cost guidance (higher sample rate and more channels cost memory/performance; Brute Force vs PCAKDTree search; pruning to save memory) and no minimum or maximum amount of animation data or memory budget.
• articleWhat does Epic's Live Link Face documentation say the app outputs (live data, recorded files, timecode, OSC) at UE 5.3?
  UE 5.3's Live Link Face guide says the iOS app streams ARKit facial animation to Unreal over the network (not USB); each recording saves a .mov reference video (JPEG-compressed frames, audio, timecode) and a .csv of raw ARKit curves importable via the Beta LiveLinkFaceImporter plugin into a Level Sequence; timecode comes from the system timer, an NTP server (default time.apple.com) or Tentacle Sync, and the app accepts OSC remote control.
• articleHow does Epic's documented IK Rig Retargeter work (chains, root, retarget pose, FK/IK settings) at UE 5.1?
  The UE 5.1 retargeting page documents two IK Rigs with retarget chains, a retarget root (typically pelvis), a retarget pose for A/T-pose differences, fuzzy chain-name matching, batch retargeting via Duplicate and Retarget, and global/per-chain settings including stride warping and speed planting. It overlaps the live-docs reading in CA-B197 (C2-B21-D3-01), here pinned to the 5.1 edition.
• articleWhat does an approved animatic carry into the next stage in documented tools (scene length, panels, soundtrack, markers) and what does it not carry?
  Storyboard Pro exports each animatic scene to a Harmony scene of the same length with its panels and soundtrack portion, either as rendered reference frames or as editable drawings, 3D objects, images, videos and camera moves; transitions are dropped and markers carry time, duration and note. Storyboarder writes an ffconcat list with each board's duration in seconds, listing the last board twice for ffmpeg.
• articleWhat is an animatic and what are the first steps to time one from storyboard panels with scratch audio?
  An animatic is storyboard panels cut in sequence with a scratch soundtrack to test timing and sync; practice is to order and time panels, play, and iterate while changes are cheap. Storyboard Pro starts every panel at one second (so a new board is a basic animatic), edits durations as H:MM:SS:FF and ripples later panels by default; a Blender tutorial times slides by frame count at 25 fps and later swaps in renders of the same length.
• articleHow does Miyazaki describe developing story through storyboards while production runs, and what risks does he give?
  Miyazaki (Paris press debate reported by Midnight Eye, translated) says his films start without a finished script, the story develops as he draws storyboards while production is already under way, and calls this a dangerous way to make an animated film; Spirited Away grew to 1,415 shots from about 1,200 envisioned. AWN (2010) quotes him saying the world grows more complex as he storyboards.
• articleHow do story supervisors and directors describe building and iterating story reels, and what limits do they name?
  Pixar story supervisor Jason Katz says Pixar screens its story as reels every three to four months, reel goals vary by director, films tend to have 'two lives' with story staff rotating around screening four or five, and Coco's first version centred on death before research shifted it; Klaus director Sergio Pablos records performances, pairs them with storyboards and cuts a story reel to test how scenes play, including trying music in the reel. Practitioner accounts; reel limits not stated.
• articleWhich Lottie / After Effects features do rlottie, ThorVG, Skottie and lottie-web document or log as unsupported?
  Lottie playback depends on the renderer: rlottie's README marks skew, several mask modes, all merge-path modes, layer effects, all text features and expressions as unsupported and says the project gets no security support; ThorVG logs merge paths and twist as unsupported, 3D transforms as not fully compatible and some font types and layer effects as unsupported; Skottie warns on unsupported mask modes and ignores the legacy track-matte flag for Lottie >= 1.0. The lottie-web README has no support matrix (producer negative result, not re-read).
• articleWhat does Lasseter's 1987 paper say about timing, spacing, anticipation and follow-through when applied to 3D keyframe animation, and what 3D-specific cautions does it give?
  Lasseter (1987) maps the principles to 3D keyframing: timing conveys weight and size, anticipation must be larger or the action slower for very fast actions, loose parts overlap and drag by material (feet leaving one or two frames apart), slow in/out is set via spline tension/bias/continuity (with overshoot risk on large changes over few frames), and fast-motion strobing is solved by motion blur or, failing that, volume-preserving squash and stretch.
• articleWhat do reviews say about the limits of marker gap filling methods, including learned methods, for motion capture used in animation?
  A CC BY 2025 Sensors review reports learned mocap recovery overfits benchmarks (CMU, HDM05) and degrades on long gaps and new subjects, clothing or marker sets, while classical interpolation is simple and interpretable but relies on manual input; it cites tree-based regression for long gaps. Method-level claims are secondary.
• articleDo biomechanics sources agree on how to choose a Butterworth cutoff, and does residual analysis give one answer?
  On eight data sets, four automatic cutoff-selection methods (Challis, Yu, Winter, Jackson) gave widely different optimal cutoffs for the same trial (walking toe x: 12.5, 3.0, 10.0, 4.0 Hz); the effect depends on the parameter (minimum toe clearance varied under 1 mm), and cutoffs above 20 Hz were judged inappropriate for force-plate data. Applying these biomechanics results to animation capture is a proposal.
• articleWhat does a mocap vendor document for filling marker gaps and choosing a smoothing cutoff, including limits on how much gap to fill?
  OptiTrack Motive recommends trimming about 3 frames at gap ends, filling only gaps below a user-set Max. Gap Size (large gaps not recommended), smoothing with a 4th-degree bidirectional low-pass Butterworth at a commonly used 7-12 Hz (higher for fast, sharp motion), and retaking when a significant marker is occluded; edits change raw trajectories and overuse is discouraged.
• articleWhat does an open paper co-authored by a studio rig team say about the inverse rig problem (solving blendshape weights from mocap) and its failure modes?
  Rackovic et al. (3Lateral/Epic byline) define a good inverse-rig solution by data fidelity, stability under later hand edits, low weight cardinality, frame-to-frame continuity and respect for weight limits; they add quadratic corrective terms for realistic rigs, report SQP fits accurate meshes but with non-sparse, non-smooth weights, and claim up to 45% lower RMS mesh error with their MM algorithm on their own benchmarks.
• articleHow do Alembic point caches (OPoints/IPoints) differ from PC2/MDD in identity, topology change and time sampling, and what does a mismatch look like?
  Alembic point samples carry positions with optional uint64 ids, velocities and widths; the API comments describe ids on the first sample and re-stated ids/topology to mark a topology change, and time is per-property TimeSampling (Uniform, Cyclic or Acyclic). The writer enforces positions on sample 0 but substitutes an empty ids array, so readers cannot assume ids exist. PC2/MDD carry neither ids nor topology.
• articleWhat exact binary layouts and checks does Blender's Mesh Cache modifier apply to PC2 and MDD files (header, vertex count, frame count, endianness)?
  Blender's PC2 reader expects a 32-byte native-endian header ('POINTCACHE2' + file_version, verts_tot, start, sampling, frame_tot) and MDD a big-endian pair (frame_tot, verts_tot) followed by per-frame float times; both refuse a vertex-count mismatch or frame_tot <= 0, and neither checks vertex order, which the manual says must stay consistent.
• articleDoes Blender Rigify document any animation retargeting or humanoid normalisation, or only rig generation from a metarig?
  Rigify documents rig generation from predefined meta-rigs (scale 1 unit = 1 m; slight knee/elbow bends for direction) but no animation retargeting or humanoid normalisation; its only 'retarget' wording concerns driver rewiring.
• articleWhat does Unity require for a humanoid Avatar so animation can be shared between models, and what does the manual say happens when the mapping or pose is wrong?
  Unity Humanoid retargeting needs a model with at least 15 bones loosely matching a human skeleton, mapped to a Humanoid Avatar in a T-pose; a successful map only means required bones matched, and recovery uses Clear/Sample Bind-pose/Automap/Enforce T-Pose.
• articleWhat does Unreal IK Rig / IK Retargeter documentation require of two skeletons before humanoid animation can be retargeted, and which mismatches does it say break or degrade the result?
  Unreal retargeting needs, on both IK Rigs, a retarget root (pelvis) and named bone chains matched by fuzzy name, plus matched retarget poses (e.g. A-pose source vs T-pose target); auto-align by local rotation axes can give nonsensical results when bone rotation conventions differ.
• articleWhat subset of Lottie does the Lottie Animation Community specification 1.0.1 actually define, so an agent knows when a feature is out of spec?
  The Lottie Animation Community spec (1.0.1) is explicitly a work in progress covering a subset of features (layer types shape/image/null/solid/precomposition; basic shapes, groups, fills/strokes/gradients, trim path), so files using other After Effects features fall outside the defined spec.
• articleWhat do the OpenXR face and eye tracking extensions define about blendshape counts, timing and ratification?
  At OpenXR-SDK 1.1.65, XR_FB_face_tracking2 (ext 288, rev 1, not ratified) exposes 70 expression weights, 2 confidence areas and a visual or audio data source, each result carrying an XrTime that is tracked or extrapolated; XR_FB_eye_tracking_social (ext 203) returns per-eye gaze with time; XR_HTC_facial_tracking defines 14 eye and 37 lip expressions.
• articleWhat version rule does the Rive runtime enforce on .riv files, and what is the runtime licence?
  At rive-runtime 91a25f40 (runtime-v0.1.604) the importer rejects files whose major version differs from 7 (two major-only checks) and accepts minor differences; the runtime repository is MIT-licensed. This says nothing about the Rive editor or hosted service terms.
• articleWhat does Spine document about its JSON/binary export, version compatibility with runtimes, and runtime licensing?
  Spine's versioning page says the editor major.minor used for export must always match the Spine Runtimes version; the Spine Runtimes License (last updated April 5, 2025) ties use of the runtimes to Spine Editor licences held by each user of the Products.
• articleDoes natural speech really have the picture (mouth movement) ahead of the sound by about 150 ms, and what offset range should a generated talking-head and its separately generated voice be expected to show?
  Schwartz and Savariaux (PLoS Comput Biol 2014) dispute the "visual speech leads by 100-300 ms" claim: for chained syllables they measured audio-visual asynchrony between about 20 ms audio lead and 70 ms audio lag, and extrapolate 40 ms audio lead to 200 ms audio lag for richer speech. Large picture leads occur at utterance starts or isolated syllables (preparatory gestures). Their data come from a simple French /a/-context corpus filmed at 50 Hz, so timing precision is about 10-20 ms. Love et al. and others still cite the 100-300 ms figure as background, so the sources conflict on the natural lead.
• articleWhich of the twelve animation principles can I expect motion capture to give a 3D character, and which need manual keyframe work, according to open scholarly sources?
  The most direct open source found is Mohr's 2019 QUT PhD thesis (CC BY-NC-ND 4.0 per the repository record). It reports, from Menache (2011, a book not read here), a three-way split: squash and stretch and anticipation, follow-through and exaggeration beyond physical limits 'cannot be accomplished with mocap'; overlapping action, straight-ahead action, ease in/out, arcs and secondary motion occur naturally; timing, appeal and personality work either way. Mohr tests this by practice (four cycles with several performers, the last with a mime artist) and finds cartoon-style movement is achievable at the capture stage under stated conditions, while squash and stretch stayed bound to the performer's body and pose-to-pose capture was used to get past physical limits. Evidence is qualitative practice research, not a controlled comparison; a 50-student experiment (Mou 2018) is consistent on extreme poses but does not test the twelve principles.
• articleDoes a public-domain English translation of Stanislavski contain the 'super-objective' and 'units and objectives' vocabulary, and if not, what does it say about a role's inner action and will?
  Negative result with partial substitute. The lawful public-domain translation found here (J. J. Robbins, My Life in Art, 1924, read in a 1938 printing) does not contain the terms super-objective, objective, unit or through-line (0 hits in the whole OCR text after normalisation). It does record Stanislavski's earlier language: an 'inner line' of a play, 'inner action', 'real necessity for action' and an inner technique 'based on the process of will'. The 'Units and Objectives' and 'The Super-Objective' chapters belong to An Actor Prepares (Hapgood translation, 1936), which a lending-library scan shows but whose rights are not open; it was not quoted. Whether an openly licensed or public-domain translation of the later system books exists remains unresolved.
• articleWhich parts of the wayang kulit night structure do the sources treat as fixed and which as flexible, what is pathet as a musical-narrative device, and what are the transfer limits for contemporary large-scale pacing?
  A study of a Klaten (Jombor) Bratayuda performance finds the division into three pathet is the part of the pattern followed strictly, while the internal order of jejer, adegan and perang is flexible: a jejer may be followed by adegan but no battle, by battle only, or by scenes that continue in a different pathet; scene description may be absent; sulukan, gending and cue sounds are used as needed. A 2004-2009 Surakarta programme study found three kinds of structure: the old pattern, compact (padat) staging that ignores the pathet sequence, and structures not following the usual order. Sumarsam's paper shows pathet itself has been defined many ways (e.g. as the dalang's singing, by gendèr tuning, by melodic patterns), and that the names nem, sanga and manyura lack an agreed explanation. For a contemporary storyteller this supports adopting 'three mode-defined blocks with a repeating internal micro-structure' as a proposal only.
• articleWhat do the open sources say about the gwangdae narrator-singer's craft criteria and the tension between improvised adaptation and a fixed text, and which failure modes (relevance, freezing, over-entertaining) follow for a solo screen or audio storyteller?
  A 19th-century theorist (Sin Jaehyo, via the thesis) named four conditions for a gwangdae: appearance/presence (inmulchire), narration (saseolchire: dramatic, literary, clear to listeners), vocal quality (deoneum) and gesture with props (neoreumsae). The same thesis shows a modern competition where judges weighed original essence, creativity and entertainment, and an expert warned that a filial-sacrifice story may not move modern audiences. Yates-Lu shows the opposite pressure: preservation systems fix lineages, and scholars dispute whether a stabilised version is the 'original', with one scholar arguing differences simply reflect a narrative not yet stable. Together they frame an open tension between adapting to the audience and preserving a text; none of the sources tests either for audio or screen use.
• articleWhat scores do these papers report on HumanML3D-based benchmarks, how comparable are they, and what could an agent check before animation review?
  PhysDiff Table 1 reports on HumanML3D (mm): MDM Penetrate 11.291, Float 18.876, Skate 1.406, Phys-Err 31.572; PhysDiff w/ MDM 0.998, 2.601, 0.512, 4.111; MotionDiffuse 20.278, 6.450, 3.925, 30.652 (rows sum correctly; no mocap ground-truth row). GMD reports MDM foot-skating ratio 0.284 and 0.161-0.309 across variants in trajectory-conditioned tests. Scores come from authors' own runs on different protocols, so they show direction (physics projection reduces artefacts) not a portable threshold. HumanML3D is 20 fps, so a clip's frame rate matters for any per-frame threshold.
• articleWhat does HOOD, an open learned cloth simulator trained without a ground-truth solver, report about quality, speed and limits compared with MeshGraphNets-style and physics-based cloth?
  HOOD (CVPR 2023, arXiv v3) combines graph networks with hierarchical message passing and trains self-supervised with a physics-based loss instead of solver-generated targets. Its human study (30 participants, pairwise realism) favours HOOD over two learned baselines and is about equal to the ArcSim physical simulator on the sequences where ArcSim could run; HOOD runs at about 7-14 fps for a 12K-vertex dress ablation (not a comparison with ArcSim time). Stated limits: no garment-garment interaction (self-penetration), failure when body motion exceeds training velocities, and wrong correspondences under severe body self-intersection. Code is MIT-licensed per the LICENSE text; data and SMPL models are fetched separately.
• articleWhat do studio and vendor first-person accounts say about the real speed and accuracy gains of ML deformers in production rigs, and where do they fall short?
  DreamWorks and Berkeley (2018) report 5x-10x faster deformation evaluation on production rigs. Golaem (DigiPro 2024) tried to reproduce this in a commercial crowd product and reports only x1.2-x2 on production assets because the paper's rigs were optimised for the model and the rest of rig evaluation dominates; its main disappointment was no significant speed gain. It reports Euler-angle and facial limits of the first method, ships the ML deformer as a crowd-package feature for close-up facial deformation, and notes crowd artists prefer overfitting because crowd sims blend clips rather than generate unseen animation. The paper's own Table 1 fps values imply 1.8x-2.5x, which does not match its text range.

More