Mixamo Animations and Characters

A complete snapshot of the Mixamo library: 2,317 motion clips and 114 rigged characters, exported as binary FBX (FBX 7.7 / fbx7_2019) with per-file metadata.

All animations share one uniform 65-joint mixamorig skeleton, so any clip can drive any compatible character without remapping.

Use animation_motion/ and character_refined/. The full export contains 2,446 animation files, but 129 are single-pose assets that carry no motion (Mixamo's *_Pose* library, plus T-Pose and Leaning); animation_motion/ is the 2,317 that actually move.

Dataset structure

animation_motion/           2,317 .fbx   clips that move  <- use this
animation/                  2,446 .fbx   the full export, poses included
character/                    114 .fbx   T-pose characters with skin mesh, as exported
character_refined/            114 .fbx   same characters, bone namespaces normalised  <- use this
character_refined_glb/        114 .glb   character_refined/ converted to GLB, same stems
character_tpose/              114 .png   rest-pose render grid per character
animation_motion_ybot/      2,317 .fbx   every moving clip baked onto Y Bot — mesh + motion
                                         in one ready-to-play file
animation_motion_render/    2,317 dirs   four-view render of every moving clip on Y Bot
                                         (MP4 + camera JSON per view)

animation_motion_prompts.json            filename -> text label + Mixamo id, 2,317 entries
animation_motion_frames.json             filename -> keyframe count, 2,317 entries
static_clips.csv                         the 129 files left out, with the measurement behind it

animation_prompts.json                   the same, over all 2,446
animation_frames.json                    the same, over all 2,446
animation_bones.json                     the 65 bone names used by every animation
metadata.csv                             browsable index of all 2,446 animations
characters.csv                           browsable index of all 114 characters
characters.json                          filename -> character name + uuid
character_bones.json                     filename -> that character's bone list
character_bone_fix_applied.txt           the 27 characters whose bones were renamed
character_nonstandard_rig.txt            the 4 characters not using `mixamorig:` naming
scripts/                                 download, indexing, normalisation and validation code

Every JSON index is keyed by FBX filename and maps 1:1 onto the files. The animation_motion_* pair covers exactly the 2,317 files in animation_motion/; the unsuffixed pair covers all 2,446. Files in animation_motion/ are the same bytes as their counterparts in animation/.

Static clips

The 129 pose files were found by decoding every animation curve and measuring its peak-to-peak range, not by filename — the split is unambiguous (largest static reading 0.024, smallest moving one 1.115, nothing between). Quiet clips like breathing idles and held guards are kept. static_clips.csv lists all 129 with the measurement and the reason.

Tabular index

Two CSVs make the dataset browsable in the dataset viewer and loadable without downloading any FBX. metadata.csv — one row per animation, covering all 2,446 files:

characters.csv — one row per character, with bone_count and covers_animation_rig (true for the 97 characters whose skeleton covers all 65 animation bones).

from datasets import load_dataset

anims = load_dataset("Linzhan/Mixamo-Animations-Characters", "animations", split="train")
chars = load_dataset("Linzhan/Mixamo-Animations-Characters", "characters", split="train")

mid = anims.filter(lambda r: 30 <= r["frames"] <= 300)                 # 1–10 s at 30 fps
ok  = chars.filter(lambda r: r["covers_animation_rig"] == "true")      # drop-in retargetable

Animations

• Count: 2,317 clips that move (2,446 including poses)
• Format: binary FBX 7.7, 30 fps, skeleton + animation curves only (no mesh)
• Skeleton: identical 65-joint mixamorig rig in every clip
• Length: 8 – 1,884 keyframes in animation_motion/

Files are named after the motion, e.g. Jab_Cross.fbx; distinct motions sharing a Mixamo name get a numeric suffix (Waving.fbx, Waving_2.fbx).

animation_motion_prompts.json

"180_Turn_W_Briefcase.fbx": {
  "prompt": "180 Turn W/ Briefcase",
  "description": "180 Turn With Briefcase",
  "motion_id": "c9c9f110-b96c-11e4-a802-0aaa78deedf9"
}

For text-conditioned tasks, description or prompt is the more discriminative label (2,134 unique strings vs 1,541 from prompt alone).

animation_motion_frames.json

"180_Turn_W_Briefcase.fbx": 68 — keyframe count parsed from each FBX, not a nominal duration.

animation_bones.json

{ "mixamo": ["mixamorig:Head", "mixamorig:HeadTop_End", "mixamorig:Hips", ...] } — the 65 bone names present in every animation: hips/spine/neck/head chain, both arms with full four-joint fingers, and both legs with toes.

Animated characters (animation_motion_ybot/)

The raw animation FBXs carry no mesh, so this folder ships every clip in animation_motion/ baked onto character_refined/Y_Bot.fbx — the character whose skeleton the animations are authored on, so no retargeting touches the motion. One FBX per clip, same stems as animation_motion/: armature, skinned mesh, materials and the full-length clip in a single ready-to-play file (~3 MB each, 6.8 GB total). Drop one into Blender or any FBX viewer and press play; these are also the exact inputs the renders below were made from.

Renders (animation_motion_render/)

A four-view render of every clip in animation_motion/, driven on Y Bot (the files in animation_motion_ybot/). One folder per clip, named after the FBX stem:

animation_motion_render/Jab_Cross/
  v000.mp4  v000.json      front  (azimuth  -90°)
  v001.mp4  v001.json      back   (azimuth  +90°)
  v002.mp4  v002.json      left   (azimuth  180°)
  v003.mp4  v003.json      right  (azimuth    0°)

Rendered with Blender EEVEE at 512×512, 30 fps, elevation 0, and composed into one H.264 MP4 per view — the transparent-film frames flattened onto a light-grey background, since H.264 carries no alpha. The MP4s preview each motion directly on the Hub; the per-frame RGBA PNGs behind them are not hosted (they would add over 800k files), so extract stills from the MP4s (ffmpeg -i v000.mp4 frames/%04d.png) if you need them.

Rendering covers the first 200 frames of a clip (~6.7 s); 375 of the 2,317 clips are longer and are truncated — frame_start/frame_end in the per-view JSON record the covered range. Frame counts otherwise match animation_motion_frames.json exactly.

Each v00X.json holds the camera in NeRF convention: view, azimuth/elevation (radians), resolution, camera_angle_x (horizontal FoV, radians), the 4×4 camera-to-world transform_matrix, the scene_scale/scene_offset that normalised the character into frame, and frame_start/frame_end/fps.

Characters

• Count: 114 characters, T-pose, with skin mesh and materials
• Rig size: 37 – 112 bones (Mixamo characters genuinely differ; some omit finger joints, others add eyes or extra joints)

characters.json maps each file to its Mixamo name and uuid; character_bones.json lists the actual bone names per character.

character/ vs character_refined/

27 of the 114 characters export with a numbered bone namespace — mixamorig1:Hips instead of mixamorig:Hips — which does not match the animations, so retargeting silently fails. character/ keeps the files exactly as exported; character_refined/ renames the namespace back to mixamorig: with a surgical binary-FBX rewrite that leaves every non-string property byte-identical (verified both ways by scripts/verify_refined.py). Use character_refined/. The affected files are listed in character_bone_fix_applied.txt.

character_refined_glb/

The same 114 characters converted one-to-one to GLB — mesh, rig and materials with textures embedded, no animation — for toolchains that read glTF rather than FBX. Same stems as the FBXs.

character_tpose/

One 1024×1024 PNG per character, named after the FBX stem: a 2×2 grid of its rest pose — front and back on the top row, left and right below — rendered (EEVEE) from character_refined/. A visual index of the 114 characters.

Retargeting compatibility

Processing it into motion data

This repository is the raw layer. The pipeline that turns it into canonicalized, text-paired motion clips is open at UniMate/data_process: it exports each clip to NPZ, renders previews, captions them with a vision-language model, cleans the joint labels and derives the training features. The processed release built from it is UniML3D.

bash data_process/scripts/run_download.sh mixamo
bash data_process/scripts/run_export.sh mixamo

Usage

import json

prompts = json.load(open("animation_motion_prompts.json"))
frames  = json.load(open("animation_motion_frames.json"))

# text label for a motion-conditioned model
text = {f: (v.get("description") or v["prompt"]) for f, v in prompts.items()}

# clips between 60 and 300 frames
clips = [f for f, n in frames.items() if 60 <= n <= 300]

Load an animation onto a character in Blender:

import bpy
bpy.ops.import_scene.fbx(filepath="character_refined/Adam.fbx")
bpy.ops.import_scene.fbx(filepath="animation_motion/Jab_Cross.fbx")
# both armatures now share the `mixamorig:` bone namespace

Known limitations

• 62 clips have no description; 23 of those share a prompt with other clips and cannot be told apart from text alone. Even preferring description, 341 clips share a label with another clip.
• In animation_motion_prompts.json, Female/Male style prefixes and leading Sequence - markers have been stripped from prompt/description — they describe animation style, not motion semantics. animation_prompts.json, metadata.csv and the filenames keep the original Mixamo labels.
• Mixamo MotionPack bundles are excluded, as their constituent clips are present individually.

Provenance

Exported from Mixamo through its own export API (format=fbx7_2019, skin=false for animations, skin=true for characters), retargeted onto the standard Y Bot skeleton. Every FBX parses without truncation, every index key matches its file, and all names/ids were cross-checked against the Mixamo API; scripts/ reproduces the download, the indexes and all checks.

Terms

Mixamo content is provided by Adobe and remains subject to Adobe's Mixamo terms. This repository is a convenience mirror for research use; consult those terms before redistributing or using the assets in a product.

Citation

This dataset is part of UniML3D, the training corpus introduced in UniMate: One Unified Model to Animate Diverse Skeletons (SIGGRAPH Asia 2026; paper page). If you use it, please cite the paper:

@article{mou2026unimate,
  title   = {UniMate: One Unified Model to Animate Diverse Skeletons},
  author  = {Mou, Linzhan and Lei, Jiahui and Dou, Zhiyang and Cai, Chenyue and Song, Chaoyue and Finkelstein, Adam and Rusinkiewicz, Szymon},
  journal = {arXiv preprint arXiv:2609.05415},
  year    = {2026}
}

Please also cite this dataset repository:

@misc{mixamo_animations_characters,
  title  = {Mixamo Animations and Characters},
  author = {Linzhan Mou},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Linzhan/Mixamo-Animations-Characters},
  note   = {FBX motion clips and rigged characters exported from Adobe Mixamo}
}