Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
JavaScript 3d-generationai-agentsanimationavatardigital-humangenerative-aiglbgltfimage-to-3dllmmcpmcp-servermodel-context-protocolsolanatext-to-3dthreejsweb-componentswebglwebxrx402
docs: drop third-party avatar vendor brands from seven published articles
nirholas0cf94cb6b202
diff --git a/docs/aws-builder-center-universal-retargeting.md b/docs/aws-builder-center-universal-retargeting.mdindex 91c568901..1ed791ef9 100644--- a/docs/aws-builder-center-universal-retargeting.md+++ b/docs/aws-builder-center-universal-retargeting.md@@ -63,7 +63,7 @@ A library tied to one spelling animates one family of avatars. We wanted it to a## 2. One canonical skeleton, and why fingers decide everything-The clips are produced by a build script ([`scripts/build-animations.mjs`](https://github.com/nirholas/three.ws/blob/main/scripts/build-animations.mjs)) that retargets source motion onto one reference avatar, an Avaturn-rigged model checked in as `public/avatars/cz.glb`, and writes each clip out as three.js JSON. Every track in the library therefore addresses one of 52 canonical joint names:+The clips are produced by a build script ([`scripts/build-animations.mjs`](https://github.com/nirholas/three.ws/blob/main/scripts/build-animations.mjs)) that retargets source motion onto one reference avatar, a photo-avatar rig checked in as `public/avatars/cz.glb`, and writes each clip out as three.js JSON. Every track in the library therefore addresses one of 52 canonical joint names:```jsexport const CANONICAL_BONES = Object.freeze([@@ -88,7 +88,7 @@ export const MIN_COVERAGE = 0.5;`canonicalizeBoneName(name)` takes any string and returns a canonical name or `null`. It works in stages, and the order of the stages is where the subtlety lives.-**Strip namespaces and vendor prefixes.** A leading namespace (`Character1:Hips` from Autodesk HumanIK, `subject:LeftUpLeg` from mocap tools) goes first. Then known vendor prefixes, each one a single line: `mixamorig:` and its numbered variants, Blender's `Armature_`, Rigify's `DEF-`/`ORG-`/`MCH-` layers, CharacterStudio's `CH_`, Reallusion's `CC_Base_`, Sketchfab's `j_`, the 3ds Max `Bip01 ` prefix, and the numbered `5.joint_` form a VRM converter emits.+**Strip namespaces and vendor prefixes.** A leading namespace (`Character1:Hips` from Autodesk HumanIK, `subject:LeftUpLeg` from mocap tools) goes first. Then known vendor prefixes, each one a single line: `mixamorig:` and its numbered variants, Blender's `Armature_`, Rigify's `DEF-`/`ORG-`/`MCH-` layers, the `CH_` prefix some avatar builders emit, Reallusion's `CC_Base_`, Sketchfab's `j_`, the 3ds Max `Bip01 ` prefix, and the numbered `5.joint_` form a VRM converter emits.**Collapse separators and case, then look up.** What remains is folded into one key, and the key is looked up in three tables in a fixed priority order:diff --git a/docs/coinmarketcap-article-play.md b/docs/coinmarketcap-article-play.mdindex 4555d7b41..0a692290b 100644--- a/docs/coinmarketcap-article-play.md+++ b/docs/coinmarketcap-article-play.md@@ -229,7 +229,7 @@ You choose how you look in every world, and there are many ways to do it.**Change without leaving.** Inside any world, press V or tap the Avatar button to swap your look on the spot: choose from your saved avatars, community quick picks, the full public gallery, a pasted link, an uploaded `.glb` or `.vrm` file, or the avatar creator. Your new look appears on you instantly, and everyone else sees it live.-**Every rig is welcome.** three.ws reads the bone names of whatever avatar you bring (Mixamo, Avaturn, Unreal, VRM and VRoid, VRM 1.0, Daz and Genesis, MakeHuman, Blender's naming style, and more) and maps them onto one shared skeleton, so an avatar made almost anywhere walks, idles, and emotes correctly the moment it loads. Compressed models load smoothly too, including Draco- and meshopt-compressed files.+**Every rig is welcome.** three.ws reads the bone names of whatever avatar you bring (Mixamo, photo-avatar rigs, Unreal, VRM and VRoid, VRM 1.0, Daz and Genesis, MakeHuman, Blender's naming style, and more) and maps them onto one shared skeleton, so an avatar made almost anywhere walks, idles, and emotes correctly the moment it loads. Compressed models load smoothly too, including Draco- and meshopt-compressed files.## Photo mode, zen mode, and sharingdiff --git a/docs/huggingface-agent-feedback-loop.md b/docs/huggingface-agent-feedback-loop.mdindex ee80ba3c3..6a6a472d2 100644--- a/docs/huggingface-agent-feedback-loop.md+++ b/docs/huggingface-agent-feedback-loop.md@@ -107,7 +107,7 @@ Together they make a refine loop you can evaluate: perception says whether it loThe other half of agentic 3D is that the asset should *do* something, and animation is where a universal approach pays off most.-We never see a model before it arrives. It might come from our own lanes, from Mixamo, VRoid, Daz, Unreal, or a Blender export from years ago. So instead of a curated allowlist of supported rigs, we canonicalise bone names, then retarget onto the canonical set. The mapping covers Mixamo, Avaturn, VRM 0.x and VRoid, VRM 1.0, Daz/Genesis, MakeHuman, the Unreal mannequin, HumanIK/Maya, Blender `.L`/`.R`, and simple `shoulderL` conventions. A new convention is one mapping entry plus a test. Rest-pose skew and hip up-axis differences are corrected automatically, every result reports its bone coverage, and a model that cannot be skeleton-driven falls back to a default rig, so every user sees a character in motion.+We never see a model before it arrives. It might come from our own lanes, from Mixamo, VRoid, Daz, Unreal, or a Blender export from years ago. So instead of a curated allowlist of supported rigs, we canonicalise bone names, then retarget onto the canonical set. The mapping covers Mixamo, photo-avatar rigs, VRM 0.x and VRoid, VRM 1.0, Daz/Genesis, MakeHuman, the Unreal mannequin, HumanIK/Maya, Blender `.L`/`.R`, and simple `shoulderL` conventions. A new convention is one mapping entry plus a test. Rest-pose skew and hip up-axis differences are corrected automatically, every result reports its bone coverage, and a model that cannot be skeleton-driven falls back to a default rig, so every user sees a character in motion.It is published on its own as `@three-ws/retarget`, with a crossfading runtime that layers one-shot gestures over a base loop.diff --git a/docs/ibm-community-any-skeleton-animation-post.md b/docs/ibm-community-any-skeleton-animation-post.mdindex fd60e1a55..ea31664ea 100644--- a/docs/ibm-community-any-skeleton-animation-post.md+++ b/docs/ibm-community-any-skeleton-animation-post.md@@ -74,7 +74,7 @@ Most write-ups solve the first and gesture at the other two, yet the second causThe foundation is boring on purpose: **pick one skeleton, author everything against it, and treat every other rig as a translation problem.**-Our canonical set is 52 joints (`CANONICAL_BONES` in [`src/glb-canonicalize.js`](https://github.com/nirholas/three.ws/blob/main/src/glb-canonicalize.js)): six for spine and head (`Hips`, `Spine`, `Spine1`, `Spine2`, `Neck`, `Head`), and per side a clavicle (`LeftShoulder`), `LeftArm`, `LeftForeArm`, `LeftHand`, fifteen finger joints, and four leg joints (`LeftUpLeg`, `LeftLeg`, `LeftFoot`, `LeftToeBase`). The reference body the clips are baked on is an Avaturn-rigged avatar, public at `https://three.ws/avatars/cz.glb`. Measured from the file, both upper arms point exactly horizontal, so it rests in a **T-pose**, with its hips 0.984 m above the floor.+Our canonical set is 52 joints (`CANONICAL_BONES` in [`src/glb-canonicalize.js`](https://github.com/nirholas/three.ws/blob/main/src/glb-canonicalize.js)): six for spine and head (`Hips`, `Spine`, `Spine1`, `Spine2`, `Neck`, `Head`), and per side a clavicle (`LeftShoulder`), `LeftArm`, `LeftForeArm`, `LeftHand`, fifteen finger joints, and four leg joints (`LeftUpLeg`, `LeftLeg`, `LeftFoot`, `LeftToeBase`). The reference body the clips are baked on is a photo-avatar rig, public at `https://three.ws/avatars/cz.glb`. Measured from the file, both upper arms point exactly horizontal, so it rests in a **T-pose**, with its hips 0.984 m above the floor.Two consequences recur throughout:@@ -153,7 +153,7 @@ Suppose every track now finds a joint. The obvious next step is to rename tracks| `michelle` (Mixamo, T-pose) | 1.0 | **body tipped 87.6 degrees, lying down** || `xbot` (Mixamo, T-pose) | 0.0 | 1.4 / -6.1, **arms stay out** || `parametric-base` (MakeHuman-derived, A-pose) | 50.4 | -96.3 / -82.9, **arms overhead** |-| `selfie-girl` (Avaturn, A-pose) | 54.7 | 72.2 / 76.9, correct |+| `selfie-girl` (photo avatar, A-pose) | 54.7 | 72.2 / 76.9, correct |`michelle` is a Mixamo export with 90 degrees baked onto the armature and the opposite on the hips; writing the clip's absolute hip rotation deletes the counter-rotation and the body tips over (our codebase still calls it "the lying-down bug"). `xbot` shares the reference's stance but not its local axes. `selfie-girl` shares the reference's axis conventions, so the copy happens to work. Hold on to that row.@@ -316,7 +316,7 @@ For a humanoid under the joint floor, the verdict from our browser-side checker**Spelling cases.** `tests/glb-canonicalize.test.js` holds 477 cases, and the six retargeting suites 649 (as of 2026-10-08). The control-rig exclusion has a test that fails if someone "fixes" it.-**A sweep that measures motion, not mapping.** A rig can map its hips and still animate as a torso with four frozen sticks, and every unit test stays green. `scripts/animation-dignity-sweep.mjs` renames one synthetic skeleton ten ways (Mixamo, Avaturn, Unreal, VRM 0.x / VRoid, VRM 1.0, Daz / Genesis 8, MakeHuman, Blender / Rigify, the simple `shoulderL` rig, anatomical Latin), so any difference is naming alone. It drives the real idle and walk through both lanes and measures limb swing and end-effector **world travel**, which catches a name-mapped bone whose parent chain broke. Run today: 10 of 10 conventions move both arms and both legs on both lanes.+**A sweep that measures motion, not mapping.** A rig can map its hips and still animate as a torso with four frozen sticks, and every unit test stays green. `scripts/animation-dignity-sweep.mjs` renames one synthetic skeleton ten ways (Mixamo, photo avatar, Unreal, VRM 0.x / VRoid, VRM 1.0, Daz / Genesis 8, MakeHuman, Blender / Rigify, the simple `shoulderL` rig, anatomical Latin), so any difference is naming alone. It drives the real idle and walk through both lanes and measures limb swing and end-effector **world travel**, which catches a name-mapped bone whose parent chain broke. Run today: 10 of 10 conventions move both arms and both legs on both lanes.**An audit against real avatars.** `scripts/audit-rig-coverage.mjs` sweeps stored avatars cheaply with a trick worth borrowing: a GLB's JSON chunk always starts at byte 20, so an HTTP `Range` request fetches the skeleton without the mesh. Its first run found five broken conventions, and fixing them moved avatars animating with a complete leg chain from 93.0% to 95.9% for uploads and 93.8% to 97.9% for imports, over 400 avatars per lane (2026-07-29).diff --git a/docs/nvidia-forum-browser-digital-human.md b/docs/nvidia-forum-browser-digital-human.mdindex f668f468a..2f8cb3a4b 100644--- a/docs/nvidia-forum-browser-digital-human.md+++ b/docs/nvidia-forum-browser-digital-human.md@@ -57,7 +57,7 @@ Our avatars come from a text prompt, a photo, or a GLB the visitor uploaded fromWe had already solved the same shape of problem for skeletons: src/glb-canonicalize.js maps Mixamo, Avaturn, Unreal, VRM, Daz, MakeHuman and plain `.L` bone conventions onto one canonical rig so the animation library retargets onto anything humanoid. The face version lives in src/voice/arkit-blendshapes.js and src/voice/a2f-player.js, and it runs in two modes per mesh:-**Direct.** The morph name canonicalizes to an ARKit shape (`jawOpen`, `JawOpen`, `jaw_open`, and the Ready Player Me, Avaturn, and MetaHuman spellings all land on the same canonical name). The A2F weight is written straight to it.+**Direct.** The morph name canonicalizes to an ARKit shape (`jawOpen`, `JawOpen`, `jaw_open`, and the web-avatar, photo-avatar, and MetaHuman spellings all land on the same canonical name). The A2F weight is written straight to it.**Derived.** The mesh only has monolithic expressions: VRM's `Aa`, `Ih`, `Ou`, `Ee`, `Oh`, or Oculus's `viseme_PP`, `viseme_FF`, `viseme_CH`. We hold a forward map from each of those to its ARKit components (`VRM_TO_ARKIT`, `OCULUS_TO_ARKIT`), and at playback we run that map **backwards**: an expression's activation is the normalized sum of the ARKit weights it is made of, taken from the A2F frame. A rig that only knows five vowels still lip-syncs, because the vowel is reconstructed out of the ARKit frame rather than looked up from a phoneme we never received.diff --git a/docs/nvidia-gtc-2027-poster.md b/docs/nvidia-gtc-2027-poster.mdindex 095b1a64d..4311647ae 100644--- a/docs/nvidia-gtc-2027-poster.md+++ b/docs/nvidia-gtc-2027-poster.md@@ -79,10 +79,10 @@ Coverage by source, from the same run (counts of distinct files):| Auto-rig | 19 | 0 | 0 | 0 || Import | 8 | 4 | 1 | 47 || Upload | 4 | 0 | 1 | 55 |-| Avaturn | 0 | 0 | 0 | 59 |+| Photo-avatar exports | 0 | 0 | 0 | 59 || Forge (text to 3D props) | 0 | 0 | 0 | 59 |-Say plainly in the poster: forge outputs are mostly props and have no face, and the Avaturn+Say plainly in the poster: forge outputs are mostly props and have no face, and the photo-avatarexports stored on three.ws carry no morph targets at all (verified on the raw files), so thoseavatars use the amplitude fallback. The 98.0% figure is coverage of rigs that have a face todrive, and the poster must label it that way.diff --git a/docs/threejs-forum-post.md b/docs/threejs-forum-post.mdindex 27db54a5a..2d12a3ae6 100644--- a/docs/threejs-forum-post.md+++ b/docs/threejs-forum-post.md@@ -30,7 +30,7 @@ Making that work produced five small, focused tools. We published each one on it**Package: `@three-ws/retarget`**-Rather than keeping a curated list of supported rigs, we canonicalise bone names first, then retarget onto the canonical set. The mapping understands Mixamo (`mixamorig:LeftArm`), Avaturn, VRM 0.x and VRoid (`J_Bip_L_UpperArm`), VRM 1.0, Daz/Genesis, MakeHuman, the Unreal mannequin, HumanIK/Maya, Blender `.L`/`.R` suffixes, and simple `shoulderL` style names. Adding a new convention is one mapping entry plus a test case, never a new code path.+Rather than keeping a curated list of supported rigs, we canonicalise bone names first, then retarget onto the canonical set. The mapping understands Mixamo (`mixamorig:LeftArm`), photo-avatar rigs, VRM 0.x and VRoid (`J_Bip_L_UpperArm`), VRM 1.0, Daz/Genesis, MakeHuman, the Unreal mannequin, HumanIK/Maya, Blender `.L`/`.R` suffixes, and simple `shoulderL` style names. Adding a new convention is one mapping entry plus a test case, never a new code path.```bashnpm install @three-ws/retarget three