TaiChi: Bridging Datasets, Graphics, LLMs & Astrophysics

Updated 7 July 2026

TaiChi is a multidisciplinary research label with diverse definitions ranging from fine-grained action datasets in computer vision to frameworks in graphics, LLM serving, and astrophysical simulations.
Its varied implementations include benchmark datasets for human motion analysis, a generative modeling benchmark (TaiChi-256), and high-performance libraries for graphics and simulation.
Researchers must carefully resolve TaiChi’s contextual meaning across disciplines to accurately apply methodologies and advance domain-specific innovations.

TaiChi, sometimes capitalized as Taichi, is a polysemous term in contemporary arXiv literature rather than the name of a single method, dataset, or software artifact. In the cited works it denotes fine-grained Tai Chi action datasets and performance-analysis pipelines; the challenging TaiChi-256 benchmark used for human image synthesis; Taichi, an open-source computer graphics library for research and prototyping; TaiChi, an LLM serving system that unifies prefill-decode aggregation and disaggregation; TaiChi, a vision-LLM framework for token communications; and Taichi, a Fast Multipole Method code for collisional stellar dynamics (Hu, 2018, Wang et al., 4 Aug 2025, Jiang et al., 28 Feb 2026, Mukherjee et al., 2020). The term therefore has to be resolved by disciplinary context.

1. Naming, scope, and disciplinary usage

The literature uses the same or nearly the same label for unrelated technical objects. In graphics systems, "Taichi: An Open-Source Computer Graphics Library" defines Taichi as a reusable infrastructure "tailored for computer graphics" (Hu, 2018). In LLM systems, "Prefill-Decode Aggregation or Disaggregation? Unifying Both for Goodput-Optimized LLM Serving" defines TaiChi as a serving architecture with "prefill-heavy" and "decode-heavy" GPU instances and three configurable sliders (Wang et al., 4 Aug 2025). In multimodal communications, "TokenCom: Vision-LLM for Multimodal and Multitask Token Communications" defines TaiChi as a VLM framework with dual-visual tokenization, a Bilateral Attention Network, and a KAN-based modality projector (Jiang et al., 28 Feb 2026). In computational astrophysics, "Fast Multipole Methods for $N$ -body Simulations of Collisional Star Systems" uses Taichi as the name of an optimized FMM code (Mukherjee et al., 2020).

Referent	Domain	Representative description
TaiChi / Tai Chi datasets	Vision, action analysis	Fine-grained Tai Chi recognition and multi-view capture
TaiChi-256	Generative modeling	Human image synthesis benchmark
Taichi library	Computer graphics	Open-source C++/Python infrastructure
TaiChi serving system	LLM inference systems	Unified PD aggregation/disaggregation
TaiChi VLM	Semantic communications	Token communication framework
Taichi code	Computational astrophysics	Optimized FMM for collisional $N$ -body dynamics

A common source of confusion is that these referents are not variants of one software lineage. The name is shared across distinct subfields, and the surrounding terminology usually disambiguates it immediately: StyleGAN3 and FID indicate the generative benchmark; pybind11, TBB, and plugin registration indicate the graphics library; TTFT and TPOT indicate the LLM serving system; and M2L, solid harmonics, and cell opening indicate the astrophysical code.

2. TaiChi as a family of human-motion datasets and benchmarks

Within computer vision, Tai Chi appears as multiple datasets with different sensing modalities, scales, and tasks. One paper studies a "small-scale fine-grained Tai Chi action dataset" with 10 classes and 20 samples per class, for a total of 200 samples, represented as skeleton sequences with $T=64$ frames, $J=25$ joints, and $C=3$ coordinates after preprocessing (Yuan et al., 2022). Another paper constructs a large multi-view RGB corpus for professional TaiChi using a 32-camera system and reports a dataset of 23 232 action clips collected from eleven practitioners performing the 24-form TaiChi sequence, with manual segmentation, action labels, and expert-assigned quality scores (Li et al., 2023). A third line of work uses the TaiChi video corpus as a source of still human images for generative modeling, treating every frame in the corpus of approximately 3 000 videos as an independent still image and resizing all frames to $256 \times 256$ for TaiChi-256 experiments (Yang et al., 2023).

These datasets support materially different problem formulations. The skeleton-based dataset targets few-shot fine-grained classification; the multi-view RGB corpus targets motion capture, surface reconstruction, and normalized performance analysis; and TaiChi-256 targets unconditional image synthesis under severe non-rigid pose variation. A plausible implication is that the recurrent adoption of Tai Chi motion reflects its combination of articulated limb trajectories, body rotation, and subtle pose differences, all of which stress geometric modeling more strongly than many daily-action benchmarks.

3. Recognition, capture, and movement-quality analysis

For fine-grained skeleton-based recognition, the paper "Spatial Transformer Network with Transfer Learning for Small-scale Fine-grained Skeleton-based Tai Chi Action Recognition" uses a spatial transformer module to learn a global affine alignment of raw 3D skeletons. The per-joint, per-frame transformation is written as

$A_\theta(x)=\theta \cdot x+b,$

with $\theta \in \mathbb{R}^{3 \times 3}$ and $b \in \mathbb{R}^3$ , followed by a Transformer-based backbone with nine hierarchical layers, learnable spatial and temporal positional encodings, multi-head self-attention with $N_h=8$ , and a temporal convolution after each spatial transformer block (Yuan et al., 2022). The transfer-learning protocol pre-trains on the NTU RGB+D Cross-View split using 49 single-actor classes, 30,349 training samples, and 15,115 testing samples, then freezes all network weights except the final fully connected classifier during fine-tuning on the Tai Chi dataset. Reported Tai Chi accuracy reaches 98.33% at the 70% training ratio with 4× augmentation, while even the 10% setting exceeds 85% accuracy (Yuan et al., 2022).

For markerless capture and performance analysis, "TaiChi Action Capture and Performance Analysis with Multi-view RGB Cameras" builds a regular 16-sided capture ring, 4.5 m in diameter and 2.5 m tall, using thirty-two FLIR RGB cameras mounted on 16 vertical columns at 1 m and 2 m heights. Intrinsic and extrinsic calibration begins with Zhang’s closed-form method and is refined by global bundle adjustment minimizing total reprojection error, with typical post-optimization error of approximately 0.5 pixel (Li et al., 2023). OpenPose extracts Body-25 2D joints from synchronized views, and multi-view triangulation recovers 3D joints by DLT; optionally, skeleton bundle adjustment refines the fused pose sequence. The reconstructed sparse skeleton sequence is then coupled with dense surface reconstruction by Instant-NeRF, using COLMAP pose estimation, multiresolution hash encoding, and importance reweighting near a coarse mesh (Li et al., 2023).

The same pipeline performs normalization modeling through motion transfer. Using TransMoMo, it disentangles motion, structure, and view codes from 2D skeleton sequences and reconstructs 3D skeletons with losses $N$ 0, $N$ 1, $N$ 2, $N$ 3, and $N$ 4. After retargeting all subjects to the same canonical "coach" skeleton, the framework compares joint trajectories and instantaneous joint angles,

$N$ 5

to localize deviations in timing and posture (Li et al., 2023). In evaluation against a Perception Neuron Studio IMU rig, the reported average joint-angle and position errors are $N$ 6 and $N$ 7 cm at 30 fps, and $N$ 8 and $N$ 9 cm at 60 fps (Li et al., 2023).

A common misconception is to treat Tai Chi analysis here as standard action recognition only. The cited literature instead spans classification, sparse 3D fusion, dense radiance-field reconstruction, and normalized coach-versus-student comparison.

4. TaiChi-256 in generative modeling of human images

In generative modeling, TaiChi functions primarily as a stress test for geometry-sensitive human synthesis. "Learning Modulated Transformation in GANs" argues that style-based generators already benefit from style modulation for cross-instance variation, but regular convolution still uses fixed receptive-field locations and therefore remains limited for large, non-rigid geometric variation. For an input feature map $T=64$ 0 and spatial location $T=64$ 1, standard convolution is written as

$T=64$ 2

with predefined offsets $T=64$ 3 (Yang et al., 2023).

The proposed modulated transformation module (MTM) predicts latent-controlled spatial offsets

$T=64$ 4

forming a variable sampling grid

$T=64$ 5

with bilinear interpolation at fractional positions. The resulting MTM convolution becomes

$T=64$ 6

so that the receptive field morphs per instance under latent control (Yang et al., 2023). In StyleGAN3 integration, the same layer style code modulates both the small offset predictor and the deformable convolution. An ablation on TaiChi/ImageNet reports that inserting MTM only into the low-resolution layers from $T=64$ 7 to $T=64$ 8 captures most geometric variation without a large speed penalty, whereas extending MTM to mid- and high-resolution layers yields diminishing returns and can induce optimizer instability (Yang et al., 2023).

Training on TaiChi-256 preserves StyleGAN3’s adversarial objective and hyperparameters: non-saturating logistic GAN loss, $T=64$ 9 gradient penalty on real samples, Adam with $J=25$ 0, $J=25$ 1, $J=25$ 2, learning rate $J=25$ 3 for both generator and discriminator, batch size $J=25$ 4, and style-mixing probability $J=25$ 5 (Yang et al., 2023). No additional geometric or reconstruction losses are introduced. On this benchmark, the reported improvements are substantial: FID $J=25$ 6 decreases from 21.36 to 13.60, CLIP-FD decreases from 25.97 to 19.27, and sFID decreases from 4.79 to 3.57 (Yang et al., 2023).

Qualitatively, the baseline StyleGAN3 often produces "stiff or 'blob-like' limbs," whereas MTM allows the coarse feature map to warp toward diverse poses; the paper specifically reports that wrists and ankles collapse less often, joints remain coherent, and posture fidelity improves. An offset-visualization experiment zeroes the learned offsets at inference and causes the generator to revert to nearly identical blob-like silhouettes across latent codes, indicating that MTM is the component responsible for encoding pose variation on TaiChi (Yang et al., 2023).

5. Taichi as a graphics and simulation infrastructure

Outside Tai Chi motion analysis, Taichi is also the name of a domain-specific C++/Python library for graphics research. "Taichi: An Open-Source Computer Graphics Library" describes four design pillars: accessibility, portability, extensibility, and high performance. Accessibility is pursued through minimal external dependencies and a Python-based installer; portability through OS abstraction layers across Linux, macOS, and Windows; extensibility through plugin registration and factory-based discovery by name string such as "APIC2D" or "MLSMPM"; and high performance through a SIMD-friendly linear algebra library, TBB-based multithreading, scoped profilers, and zero-cost C++ abstractions (Hu, 2018).

Architecturally, the library separates a Python frontend, bound through pybind11, from a C++ kernel layer containing data structures, compute kernels, task scheduling, and utilities. The details explicitly list Field<T,N>, ParticleSystem, mesh and scene graph types, TBB task groups and parallel_for, serialization, logging, profiling, and wrappers for stb_image/png/tiff, tinyobjloader, and ffmpeg. Plugin registration uses macros such as TAICHI_REGISTER_PLUGIN, enabling runtime selection by string without build-script changes (Hu, 2018). The reported performance advantage of Taichi’s 16-byte-aligned 3D vectors over Eigen’s 12-byte-packed Vec3 is a 1.8×–2.3× speedup on 3D vector-matrix multiplies, and the built-in scoped profiler adds under 5% overhead when enabled (Hu, 2018).

The library’s research role is further illustrated by downstream simulation work. The SPH framework described in "Journey into SPH Simulation: A Comprehensive Framework and Showcase" uses Taichi fields for positions, velocities, densities, pressures, and temporary buffers, and expresses density summation, pressure acceleration, rigid-fluid coupling, and viscosity operators as @ti.kernel loops (Huang et al., 2024). The framework states that Taichi compiles kernels to GPU code or multithreaded CPU code and can interoperate with custom CUDA by sharing device pointers via Taichi’s low-level API and synchronizing with ti.cuda.synchronize() (Huang et al., 2024). On DFSPH with 1.23 M particles, the reported average time per step is 2.06 s on an Intel Xeon Gold 5218R CPU, 0.440 s on an AMD EPYC 7713 CPU, 0.111 s with the Vulkan backend on an RTX 3090, 0.0477 s with the CUDA backend on an RTX 3090, 0.101 s with Vulkan on an A100 Tensor Core, and 0.0401 s with CUDA on an A100 Tensor Core (Huang et al., 2024).

A plausible implication is that Taichi’s importance lies less in a single numerical method than in a systems pattern: Python orchestration, low-level kernel execution, backend portability, and plugin-level extensibility.

6. TaiChi in LLM serving and multimodal communication

In LLM inference, TaiChi denotes a serving system for balancing prefill and decode latency objectives. The paper defines Time-to-First-Token (TTFT) as the latency from request arrival until the first output token is emitted, and Time-per-Output-Token (TPOT) as the average time for subsequent autoregressive decode tokens. It argues that conventional PD aggregation excels under tight TTFT and relaxed TPOT, while PD disaggregation excels under strict TPOT and relaxed TTFT, but that neither is optimal under balanced SLOs. TaiChi addresses this with a unified disaggregation-aggregation architecture using "prefill-heavy" and "decode-heavy" GPU instances, along with three configurable sliders: the instance ratio $J=25$ 7, the chunk size $J=25$ 8 on P-heavy GPUs, and the chunk size $J=25$ 9 on D-heavy GPUs (Wang et al., 4 Aug 2025).

The core scheduling mechanisms are flowing decode scheduling and length-aware prefill scheduling. Decode always starts on a D-heavy instance; when D-heavy memory pressure exceeds a watermark, requests with the longest current output length are migrated to P-heavy machines to release low-interference decode capacity, and requests on P-heavy machines can flow back to D-heavy when real-time TPOT approaches $C=3$ 0 (Wang et al., 4 Aug 2025). Prefill assignment estimates queueing time, execution time, and, for P-heavy instances, KV transfer time, and selects a feasible instance satisfying the TTFT constraint with the smallest queued prefill token count. On eight NVIDIA A100 80 GB GPUs with Qwen2.5-14B and Qwen2.5-32B, the paper reports up to 77% goodput improvement over the best baseline in balanced regimes; in summarization SLO1 on the 14B model, PD aggregation reaches 4.2 QPS, PD disaggregation 3.1 QPS, and TaiChi 7.4 QPS. At TaiChi’s maximum goodput, TTFT P90 is reduced by 2.4×–13.2× relative to PD disaggregation and TPOT P90 by 1.11×–1.69× relative to PD aggregation, while KV transfers plus scheduling overhead remain under 1% of total request time (Wang et al., 4 Aug 2025).

In multimodal semantic communications, TaiChi instead denotes a VLM framework. The architecture begins with a dual-visual tokenizer: a low-resolution ViT branch for global semantics and a high-resolution convolutional branch for fine local details. These token streams are fused by a Bilateral Attention Network through top-down and bottom-up localized cross-attention, then compressed into a residual aggregate

$C=3$ 1

with an ideal compression ratio

$C=3$ 2

The fused tokens are projected into text-embedding space by a KAN-based modality projector whose univariate activations are learned as a combination of SiLU and a B-spline basis, and the resulting multimodal token sequence is jointly optimized with channel coding under cross-entropy loss (Jiang et al., 28 Feb 2026). Reported zero-shot benchmark scores for TaiChi with Gemma-2B are 59.2/61.9/1323 on VQA $C=3$ 3/MMB/MME, while Qwen2.5-14B yields 65.1/73.5/1565. Ablations without KAN and without BAN degrade performance, and the token-communication system reportedly outperforms prior token-SC and DeepSC baselines by 5–10 percentage points at low SNR while retaining above 90% accuracy above 10 dB (Jiang et al., 28 Feb 2026).

These two TaiChi systems are unrelated in mechanism but similar in naming style: both define hybrid architectures that allocate limited computational resources adaptively across heterogeneous pathways.

7. Taichi in collisional $C=3$ 4-body dynamics

In computational astrophysics, Taichi is an optimized Fast Multipole implementation for collisional star systems. The code replaces $C=3$ 5 direct summation with tree-based FMM using solid spherical-harmonic multipole and local expansions, an error-controlled cell-opening criterion, and a rotation-accelerated multipole-to-local (M2L) translation operator (Mukherjee et al., 2020). For cells $C=3$ 6 and $C=3$ 7, the opening test combines a geometric condition with force-scale and error-bound estimates:

$C=3$ 8

The M2L operator is factored by rotations and coordinate swaps so that a general translation reduces to a precomputed $C=3$ 9-aligned translation,

$256 \times 256$ 0

which reduces cost from $256 \times 256$ 1 to $256 \times 256$ 2 per cell pair (Mukherjee et al., 2020).

The accuracy regime reported for $256 \times 256$ 3 is close to direct summation. For a $256 \times 256$ 4-particle Plummer model with $256 \times 256$ 5, the median fractional force error is at most $256 \times 256$ 6 and the 99.99th percentile is approximately $256 \times 256$ 7. Long integrations show cumulative $256 \times 256$ 8 up to core collapse, and Lagrangian radii agree with NBODY6++GPU to within about 1% across mass shells, with an average shell-radius difference of at most 0.1% at $256 \times 256$ 9 (Mukherjee et al., 2020). The code also reproduces collisional effects such as dynamical friction, core collapse times, and the theoretical $A_\theta(x)=\theta \cdot x+b,$ 0 density profile at collapse (Mukherjee et al., 2020).

Performance crosses over in favor of FMM beyond roughly $A_\theta(x)=\theta \cdot x+b,$ 1. At $A_\theta(x)=\theta \cdot x+b,$ 2, the FMM configuration with $A_\theta(x)=\theta \cdot x+b,$ 3 is reported as more than 100× faster than Taichi’s own direct-summation kernel and more than 10× faster than NBODY6++GPU on a 28-core Xeon; strong scaling on 1 to 28 physical cores reaches approximately 16×, with OpenMP task parallelism and AVX intrinsics accelerating the tree walk and near-field sums (Mukherjee et al., 2020). The paper identifies future work in binary and few-body regularization, higher-order integrators with jerk evaluation, refined symmetric/adaptive stepping, and hybrid long-range/neighbor-splitting schemes.

Across these disparate literatures, TaiChi is best understood not as a single canonical concept but as a context-dependent research label attached to motion datasets, generative benchmarks, simulation infrastructure, inference-serving systems, semantic-communication architectures, and astrophysical solvers.