本文へ移動
cccskills
無料GitHub で公開

scenario-unreal-animation

Use when a UE5 task involves characters in motion: "import FBX from Maya" with a skeleton, "retarget animations", IK Rig, IK Retargeter, feet sinking or floating after retarget, mocap cleanup, Motion Matching, Pose Search, Choosers, GASP, state machine vs motion matching, root motion, offset root bone, foot IK, Control Rig, Animation Blueprint cost, "animation is expensive", budget allocator, URO, MetaHuman LODs, ML Deformer, cloth.

インストール方法を見る

含まれるファイル(7)

  • SKILL.md20.8 KB
  • references/critique.md13.7 KB
  • references/expert-notes.md27.1 KB
  • references/gui-paths.md10.9 KB
  • references/procedures.md39.0 KB
  • references/sources.md13.2 KB
  • scripts/ue_anim.py147.8 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Unreal animation (technical animator)

Expert level means contact that holds on every clip the game plays, locomotion authored against the movement model that changes category on the input frame, animation that costs a measured number of milliseconds on the target, and clean data at the source. The agent works through data and templates: it scripts assets, IK Rigs, retargeters, databases and settings, duplicates Epic's AnimBPs instead of drawing graphs, measures contact numerically, then looks at frames. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

REQUIRED BACKGROUND: scenario-unreal-expert (channels, review loop, 5.8 traps).

Status (2026-09-24): Unreal is not installed. The offline layer of scripts/ue_anim.py passed 30 tests (python3 tests/code/unreal-animation/run_all.py --offline). Every in-editor snippet is not yet run in Unreal; [verify] names are answered by job_00_anim_probe.py. Import: sys.path[:0] = ["<skills>/scenario-unreal-animation/scripts", "<skills>/scenario-unreal-expert/scripts"]; import ue_anim as A.

Stance (the expert delta)

  • Cheapest transfer first, then retarget order decides contact. Same skeleton, compatible skeletons, Copy Pose From Mesh or Leader Pose beat an IK Retargeter (Greg Richardson, -8CQNaODNNA [00:02:40]-[00:06:04]). When retargeting: a new named retarget pose (never Default) with auto align, snap to ground, slight knee and elbow bend; Scale Source once at the top, judged by the IK goals; Blend to Source on leg IK or feet sink; auto-exported retargeters have IK off (Richardson [00:08:13], [00:16:29], [00:21:26], [00:29:14]). Pin the IK helper bones gameplay reads (ik_foot_l/r, ik_hand_gun); one-off fixes go in an Additive Pose op, not the base pose (Richardson [00:33:16]; 5.8 retargeting doc).
  • Movement model first, then animate against it. Tune capsule speeds, acceleration and braking with rough animation until it plays well, export the capsule trajectory, animate the root frame-exact with the chest (not the hips) within tolerance and 5-frame stops where the body overshoots. This separates animation bugs from system bugs (GASP team, Tony, mhVp_cC9MLc [00:24:25]-[00:28:43]; Dustin DeVoe, FLDXtAV7qsw [00:06:27]-[00:08:40]).
  • Motion matching is data under designer control. Split databases by movement state, pick them with a Chooser on cached enums, interrupt only when a core state changed this frame, keep root motion enabled on every locomotion clip even when the capsule drives (GASP team [00:29:43]-[00:40:54], [01:30:00]). Weights encode a preference between imperfect candidates with low coverage (Sam [00:49:01]); they never fix a speed or coverage mismatch, and late tuning uses notify-state tags, not schema weights (Jose, tNw9lD2PW3U [00:18:28]; motion matching doc).
  • Small sets ship with procedural nodes. 13 clips give a non-strafing base, 26 with strafing, 60 to 80 with states, both feet for every start, stop and refacing start (DeVoe [00:11:14]-[00:14:44]); "a misnomer that you need all coverage" (GASP [01:21:06]). Sparse data needs Steering before Offset Root Bone, foot placement and mesh-space additives, not a state machine (DeVoe [00:28:19]-[00:31:40]).
  • Keep the AnimBP off the game thread, and budget root motion. Thread-safe update, Property Access, fast path on every node (Matthew Lake, N_suMyUuork [00:12:44]-[00:16:53]). Root Motion from Everything AND from Montages Only both update the AnimGraph on the game thread (root motion doc; Lake [00:13:18]). Montages Only is chosen because capsule-driven locomotion networks well, not to save threads: budget its cost; a character with no root-motion montage can use No Root Motion Extraction [inference from the doc wording, verify in Insights].
  • Clean at the source, throttle what is unseen. Stripping twist, face and deformation tracks and Maya helper curves cut storage 71% and memory 90% on one clip; import defaults set project-wide, not per asset (Lake [00:05:15]-[00:07:21]). Only Tick Pose When Rendered as the project default, budget allocator or URO (never both), one LOD Settings asset with hysteresis (Lake [00:21:13]; AnimBP doc).
  • Control Rig: Construction caches, Forwards Solve per frame. Profile before converting to C++ (60 µs graph vs 10 to 16 µs rig unit); FBIK adjusts a pose, it does not create one; exclude bones instead of locking them; a Backwards Solve makes a rig bakeable (Stéphane Biava, XYMad1EutcA [00:10:08]; LEGO Fortnite tech animator, kO6pz8ADgcU [00:12:04]).
  • Never ship the cinematic MetaHuman. About 800 MB cooked vs about 60 MB Optimized High; audition LODs at the game camera from LOD 2 on PS5-class; Mac renders cards only (DeVoe, tTgMafRAM7A [00:15:16], [00:23:34]).

Establish first

InputChangesDefault when silent
Platform, frame budget, characters on screenbudgets, LODs, allocator, runtime vs offline retargetconsole-class 60 fps; animation share from scenario-unreal-performance
Multiplayer, hero fidelitycapsule vs root-motion-drivenmultiplayer: capsule driven, root motion in montages; single-player hero needing planted refacing turns: evaluate root-motion-driven through the root motion attribute (Jose [00:33:25]-[00:34:29]); Mover 5.8 adds ChaosMover trajectory prediction, still heading out of Experimental
Movement model: speeds per direction, acceleration, brakingtiming of every clipGASP speeds when reusing GASP data; tune before authoring
Clip set: GASP, mocap, hand-keyed, countmotion matching vs state machinemotion matching with procedural nodes; state machine plus warping (Lyra) only for explicit per-state control
Gameplay needsstrafing and states multiply content (each state almost doubles it, DeVoe)orient-to-movement, walk, run, jump
Skeleton contract from the DCCchains, IK bones, rootscenario-maya-rigging game_skeleton_check clean

Workflow

  1. Probe once per engine. job_00_anim_probe.py headless. GATE: probe.json lists IK Rig, retargeter, Pose Search, AnimPose, Interchange, MetaHuman names and whether 5.8 BlueprintGraphEditor exists.
  2. Receive and audit the character. A.import_character / A.import_animations (explicit options on either importer), then A.import_audit(A.skeletal_mesh_facts(mesh)): one root, unused bones (5.8 Bone Count Reduction), influences over 8, units. Set project-wide import defaults once: A.apply_ini_plan(cfg, A.import_defaults_plan()) for the legacy importer, the project pipeline in the Interchange Import Content stack. GATE: no error; mesh looked at in a neutral lit preview.
  3. Choose the transfer. A.transfer_method(source, target). GATE: method and reason written down.
  4. IK Rigs. A.create_ik_rig (5.8 template splits Leg at the ankle and adds Foot chains), A.biped_chain_plan for custom names. GATE: A.chain_map_check clean.
  5. Retargeter, built once, then templated. A.create_retargeter; op settings once in the GUI, saved as RTG_Template; Pin Bones from A.ik_bone_pin_plan(bones, A.mannequin_bone_map(bones, parents)). GATE: A.op_stack_check(ops, proportions_differ=True, target_ik_bones=..., runtime=..., profile_ops_us=...) clean; A.retarget_log_check on the Retarget Output Log clean; LOD thresholds and Profile Ops for any runtime retargeter.
  6. Batch retarget and clean. A.batch_retarget(..., retain_additive=True); mocap through its own retargeter (Filter Bones, Generate From Target Pelvis); A.anim_content_audit, A.ensure_root_motion, A.strip_curves. GATE: A.compare_contact on walk, run, turn, crouch, jump, a contact clip and a mocap clip; A.root_motion_check; contact frames looked at side and front.
  7. Movement model and authored clips. Tune the movement component per gait and direction; export the capsule trajectory (5.6+ Rewind Debugger > Trajectories export, else Take Recorder) to scenario-maya-animation or scenario-blender-animation. GATE: A.capsule_match (root on the path, chest within tolerance, root stops with the capsule) and A.speed_parity on returned clips.
  8. Locomotion architecture. A.decide_locomotion(...). Motion matching: migrate GASP, A.motion_matching_coverage (foot pairs, capture shapes), A.database_plan, duplicate schemas and databases, then remap every mannequin bone name: A.schema_bone_check(A.schema_facts(s)["bones"], bones, bone_map), Pose History bones in the GUI; tags, notify filtering, repetition guards. Procedural nodes: A.procedural_order_check, A.offset_root_bone_modes. GATE: zero "no database" warnings cycling every state in PIE; Rewind Debugger read on each transition frame; starts, stops, pivots at 0.25 timescale from a still camera.
  9. Root motion and montages. A.root_motion_mode_check(mode, multiplayer, montage_root_motion); A.montage_blend_out_check; traversal tails through a JustTraversed database. GATE: no root velocity in the blend-out window; show collision capsule stays with the mesh; game thread measured.
  10. Foot IK. GASP 5.4 leg IK only locks feet to the IK foot bones, no ground adaptation [verify current GASP]; slopes and stairs need Foot Placement or a Control Rig with FBIK (A.fbik_foot_defaults(mannequin_like=False) re-derives the knee axis). Sequencer shots: CR_GroundAlignment on the mannequin, then retarget. GATE: stairs and 30 degree slope captures, foot gap per frame, no pop, no jitter.
  11. Performance pass. A.anim_ini_plan, A.animbp_settings, A.compile_and_scan, A.allocator_config_check, A.lod_settings_check. GATE: stat anim, stat unit, Insights before and after on the same path.
  12. Deliver per Handoffs, with what was not verified.

Numbers

ValueRelative toSource
GASP speeds: walk about 2, strafe 3.5, run 5 m/smovement model and clipsGASP team [01:27:17]
Capsule stop 5 frames; offset root radius about 30 cmcapsuleGASP team [00:27:38], [01:52:15]
Steering desired facing 0.5 s aheadtrajectoryJose [00:29:50]
Notify Recency Time Out 0.2 s; weights example Pose 1, Trajectory 3Motion Matching node, schemamotion matching doc
Minimal sets 13, 26, 60 to 80 clips (GASP over 500)locomotionDeVoe
a.Budget.BudgetMs 1.0 (default), 1.5, 2.0, 2.5game thread, ViewDistanceQuality 0 to 3AnimBP doc
Max bone influences 8 (LOD0), 4, 1 (last)skeletal LODsLake; Biava
FBIK: pelvis stiffness 0.8, knee preferred Z 45, ankle -70..70mannequin joint axesControl Rig doc
Filter Bones: Responsiveness 0.3 to 0.8, Velocity Cutoff 15 to 30 Hz, under frame rate / 2mocapretargeting doc
MetaHuman pose drivers 1.4 ms PS5 before 5.6, about 0.4 ms afterper characterDeVoe
Contact tolerances 1 to 2 cm; root on capsule 2 cm; chest 30 cmplanted frames, authored clips[added] defaults

No source gives a per-character ms budget: take the animation share from scenario-unreal-performance.

Quality gates

  • Code: the A.* checks of each Workflow gate, summarized by verdict(issues) with no error.
  • Engine numbers: stat anim, stat unit, Insights (animation channel, game-thread AnimGraph update), stat group PoseSearch (5.8 names), a.Budget.Debug.Enabled 1, Retarget Profile Ops.
  • Visual (ue_review): contact frames side and front, native in the middle, 0.25 timescale locomotion, stairs and slopes, LOD at distance. Score with references/critique.md.

Common mistakes

MistakeWhat it looks likeFix
Animation imported without the target skeletonclip does not play; stray <Mesh>_Skeletonalways pass the skeleton
No Blend to Source on leg IKfeet sink on shorter targetsBlend to Source; Scale Source at the top if hands over-reach
IK helper bones not pinnedleg or weapon IK reads a reference-pose bonePin Bones per ik_bone_pin_plan
Batch with an auto-generated retargeterIK off, override sets ignoredhand-built retargeter
Clips authored before the movement model is tunedevery clip retimed; animation and system bugs mixedmodel first, capsule_match
Duplicated GASP schema on a custom skeletonchannels read missing mannequin bones, poor picksschema_bone_check, schema skeleton, mirror table rebuilt
One big database, weights tunedstarts improve, stops breaksplit by state, Chooser, interrupt on core state change
State machine chosen because the set is smallstiff transitionsminimal set plus procedural nodes
Steering after Offset Root Bone, or none, with keyboard inputselection thrashesSteering first, only while moving or in air
Montages Only picked "to keep the graph on workers"unbudgeted game-thread costboth modes run the graph on the game thread; budget, or No Root Motion Extraction
Blend Out Trigger Time -1 with root still movingwalk-loop frames after traversaltrigger 0, or root motion ends earlier
Allocator node without a.Budget.Enabled, or allocator plus UROno effect, confusing throttlingboth switches on; never both systems
Offset root bone near wallsmesh clips into wallssmaller radius near walls or collision-aware trajectory

Handoffs

  • Receives from scenario-maya-rigging, scenario-maya-animation, scenario-blender-rigging, scenario-blender-animation through scenario-unreal-pipeline-automation: FBX 2020.2, cm, one root at the origin, joints only, one clip per file, game_skeleton_check JSON.
  • Delivers to scenario-maya-animation or scenario-blender-animation: exported capsule trajectory, speed, acceleration and braking per direction, stop spec, chest tolerance, clip list with both feet.
  • Delivers to scenario-unreal-gameplay: character Blueprint child (GASP or Lyra), AnimBP variable contract (gait, stance, movement mode, rotation mode), montages with warp targets, movement model values, note that CMC input acceleration needs C++ replication for multiplayer trajectories.
  • Delivers to scenario-unreal-cinematics: retargeted clips, bakeable Control Rigs, CR_GroundAlignment guidance, MetaHuman Cine tier only for pre-rendered shots.
  • Asks scenario-unreal-performance for the animation budget and target captures.

UE 5.8 notes

  • Retargeter: op stack (5.6), Override Sets replace Retarget Profiles, foot definition and floor constraint, Blend to Source, Scale Goals, Offset Goals as separate Run IK Rig sub-ops; resave pre-5.8 retargeters.
  • Pose Search: DatabaseAnimationAssets (5.7), Chooser-column databases default to Brute Force, ...FromContext channel getters, renamed stats, "Block Transition In".
  • unreal.BlueprintGraphEditor is the 5.8 route for Blueprint graph edits; whether it reaches AnimGraph nodes is [verify]: try on a copy, else templates and computer use.
  • Interchange: production ready for asset import (5.5), default for FBX level import (5.6), dialog reworked (5.8); the FBX asset-import default is [verify], so pass options for both importers.
  • Control Rig: Construction Event, Physics Beta, Dynamics Experimental; FSmartName and 5.0 to 5.6 anim APIs removed.
  • MetaHuman: ConvertLegacyDNAAssets -DryRun first; face MetaHuman Animator on macOS, body Windows only.
  • Experimental, fallback required: Animation Mixer, Direct Mesh Controls, uFBX, MetaHuman Crowd, MotionMatchMulti. Mover still Experimental.

References

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Handle review comments received on GitHub for the current branch's PR or a given PR number or URL, by triaging each unresolved thread, fixing what is warranted, pushing, and replying in each thread, when asked to address, fix or answer PR comments. Works in any Scenario repository.

日本語の概要は準備中です。原文の説明を表示しています。

scenario-labs/skills9922026年10月13日 更新

Generate one short Conventional Commits squash message describing only the net diff between the PR's base and head, set its header as the PR title, validate it with the repository's commitlint and both copy its description to the clipboard and display it, when asked for a squash message. Works in any Scenario repository, whatever its default branch.

日本語の概要は準備中です。原文の説明を表示しています。

scenario-labs/skills9922026年10月13日 更新

Refresh the current pull request description from its net changes, following the repository's PR template, when asked to update or write the PR summary. Works in any Scenario repository, whatever its default branch.

日本語の概要は準備中です。原文の説明を表示しています。

scenario-labs/skills9922026年10月13日 更新

scenario

無料

Use when connecting an AI agent to Scenario (scenario.com) through MCP, or when a task involves generating images, video, 3D, audio, sprites, textures, or game assets. Also when picking a Scenario model, running a LoRA, refining a generation prompt, uploading reference images, waiting on generation jobs, checking credits or quota, hitting Scenario auth, scope, or Forbidden errors, or setting up mcp.scenario.com in Claude Code, Cursor, VSCode, or another agent.

日本語の概要は準備中です。原文の説明を表示しています。

scenario-labs/skills9922026年10月13日 更新

Use when generating or handling 3D assets through the Scenario MCP server, including text-to-3D or image-to-3D meshes, GPT-6 Astra 3D or Claude Opus 5.5 3D, GLB, FBX, OBJ, STL, or VOX files, PBR-textured or game-ready models, voxel models, multi-view reconstruction, retexture, remesh, UV unwrap, auto-rigging a biped or quadruped, or retargeting an animation, previewing a mesh in the inline 3D viewer, or downloading a model for import into Unity, Unreal, Godot, or Blender.

日本語の概要は準備中です。原文の説明を表示しています。

scenario-labs/skills9922026年10月13日 更新

Use when generating navigable 3D worlds or Gaussian splat scenes on Scenario via MCP: text-to-world, image-to-world, a 360 panorama or skybox turned into a walkable splat, multi-view photos or a video walkthrough reconstructed as a splat, or one object photo to splat. Keywords: World Labs Marble 1.1, 1.1 Plus, 1.0 Draft, Hunyuan World, HY World, image-to-skybox, skybox-to-splat, multiview-to-splat, TripoSplat, gaussian splatting, 3DGS, spz, world generation, explorable scene, VR.

日本語の概要は準備中です。原文の説明を表示しています。

scenario-labs/skills9922026年10月13日 更新

scenario-labs のスキルをすべて見る

このスキルの問題を報告する